IO tools (text, CSV, HDF5, …)#

The pandas I/O API is a set of top level reader functions accessed like pandas.read_csv() that generally return a pandas object. The corresponding writer functions are object methods that are accessed like DataFrame.to_csv(). Below is a table containing available readers and writers.

Format Type

Data Description

Reader

Writer

text

CSV

read_csv

to_csv

text

Fixed-Width Text File

read_fwf

text

JSON

read_json

to_json

text

HTML

read_html

to_html

text

LaTeX

Styler.to_latex

text

XML

read_xml

to_xml

text

Local clipboard

read_clipboard

to_clipboard

binary

MS Excel

read_excel

to_excel

binary

OpenDocument

read_excel

binary

HDF5 Format

read_hdf

to_hdf

binary

Feather Format

read_feather

to_feather

binary

Parquet Format

read_parquet

to_parquet

binary

ORC Format

read_orc

to_orc

binary

Stata

read_stata

to_stata

binary

SAS

read_sas

binary

SPSS

read_spss

binary

Python Pickle Format

read_pickle

to_pickle

SQL

SQL

read_sql

to_sql

SQL

Google BigQuery

read_gbq

to_gbq

Here is an informal performance comparison for some of these IO methods.

Note

For examples that use the StringIO class, make sure you import it with from io import StringIO for Python 3.

CSV & text files#

The workhorse function for reading text files (a.k.a. flat files) is read_csv(). See the cookbook for some advanced strategies.

Parsing options#

read_csv() accepts the following common arguments:

Basic#

filepath_or_buffervarious

Either a path to a file (a str, pathlib.Path, or py:py._path.local.LocalPath), URL (including http, ftp, and S3 locations), or any object with a read() method (such as an open file or StringIO).

sepstr, defaults to ',' for read_csv(), \t for read_table()

Delimiter to use. If sep is None, the C engine cannot automatically detect the separator, but the Python parsing engine can, meaning the latter will be used and automatically detect the separator by Python’s builtin sniffer tool, csv.Sniffer. In addition, separators longer than 1 character and different from '\s+' will be interpreted as regular expressions and will also force the use of the Python parsing engine. Note that regex delimiters are prone to ignoring quoted data. Regex example: '\\r\\t'.

delimiterstr, default None

Alternative argument name for sep.

delim_whitespaceboolean, default False

Specifies whether or not whitespace (e.g. ' ' or '\t') will be used as the delimiter. Equivalent to setting sep='\s+'. If this option is set to True, nothing should be passed in for the delimiter parameter.

Column and index locations and names#

headerint or list of ints, default 'infer'

Row number(s) to use as the column names, and the start of the data. Default behavior is to infer the column names: if no names are passed the behavior is identical to header=0 and column names are inferred from the first line of the file, if column names are passed explicitly then the behavior is identical to header=None. Explicitly pass header=0 to be able to replace existing names.

The header can be a list of ints that specify row locations for a MultiIndex on the columns e.g. [0,1,3]. Intervening rows that are not specified will be skipped (e.g. 2 in this example is skipped). Note that this parameter ignores commented lines and empty lines if skip_blank_lines=True, so header=0 denotes the first line of data rather than the first line of the file.

namesarray-like, default None

List of column names to use. If file contains no header row, then you should explicitly pass header=None. Duplicates in this list are not allowed.

index_colint, str, sequence of int / str, or False, optional, default None

Column(s) to use as the row labels of the DataFrame, either given as string name or column index. If a sequence of int / str is given, a MultiIndex is used.

Note

index_col=False can be used to force pandas to not use the first column as the index, e.g. when you have a malformed file with delimiters at the end of each line.

The default value of None instructs pandas to guess. If the number of fields in the column header row is equal to the number of fields in the body of the data file, then a default index is used. If it is larger, then the first columns are used as index so that the remaining number of fields in the body are equal to the number of fields in the header.

The first row after the header is used to determine the number of columns, which will go into the index. If the subsequent rows contain less columns than the first row, they are filled with NaN.

This can be avoided through usecols. This ensures that the columns are taken as is and the trailing data are ignored.

usecolslist-like or callable, default None

Return a subset of the columns. If list-like, all elements must either be positional (i.e. integer indices into the document columns) or strings that correspond to column names provided either by the user in names or inferred from the document header row(s). If names are given, the document header row(s) are not taken into account. For example, a valid list-like usecols parameter would be [0, 1, 2] or ['foo', 'bar', 'baz'].

Element order is ignored, so usecols=[0, 1] is the same as [1, 0]. To instantiate a DataFrame from data with element order preserved use pd.read_csv(data, usecols=['foo', 'bar'])[['foo', 'bar']] for columns in ['foo', 'bar'] order or pd.read_csv(data, usecols=['foo', 'bar'])[['bar', 'foo']] for ['bar', 'foo'] order.

If callable, the callable function will be evaluated against the column names, returning names where the callable function evaluates to True:

In [1]: import pandas as pd

In [2]: from io import StringIO

In [3]: data = "col1,col2,col3\na,b,1\na,b,2\nc,d,3"

In [4]: pd.read_csv(StringIO(data))
Out[4]: 
  col1 col2  col3
0    a    b     1
1    a    b     2
2    c    d     3

In [5]: pd.read_csv(StringIO(data), usecols=lambda x: x.upper() in ["COL1", "COL3"])
Out[5]: 
  col1  col3
0    a     1
1    a     2
2    c     3

Using this parameter results in much faster parsing time and lower memory usage when using the c engine. The Python engine loads the data first before deciding which columns to drop.

General parsing configuration#

dtypeType name or dict of column -> type, default None

Data type for data or columns. E.g. {'a': np.float64, 'b': np.int32, 'c': 'Int64'} Use str or object together with suitable na_values settings to preserve and not interpret dtype. If converters are specified, they will be applied INSTEAD of dtype conversion.

New in version 1.5.0: Support for defaultdict was added. Specify a defaultdict as input where the default determines the dtype of the columns which are not explicitly listed.

dtype_backend{“numpy_nullable”, “pyarrow”}, defaults to NumPy backed DataFrames

Which dtype_backend to use, e.g. whether a DataFrame should have NumPy arrays, nullable dtypes are used for all dtypes that have a nullable implementation when “numpy_nullable” is set, pyarrow is used for all dtypes if “pyarrow” is set.

The dtype_backends are still experimential.

New in version 2.0.

engine{'c', 'python', 'pyarrow'}

Parser engine to use. The C and pyarrow engines are faster, while the python engine is currently more feature-complete. Multithreading is currently only supported by the pyarrow engine.

New in version 1.4.0: The “pyarrow” engine was added as an experimental engine, and some features are unsupported, or may not work correctly, with this engine.

convertersdict, default None

Dict of functions for converting values in certain columns. Keys can either be integers or column labels.

true_valueslist, default None

Values to consider as True.

false_valueslist, default None

Values to consider as False.

skipinitialspaceboolean, default False

Skip spaces after delimiter.

skiprowslist-like or integer, default None

Line numbers to skip (0-indexed) or number of lines to skip (int) at the start of the file.

If callable, the callable function will be evaluated against the row indices, returning True if the row should be skipped and False otherwise:

In [6]: data = "col1,col2,col3\na,b,1\na,b,2\nc,d,3"

In [7]: pd.read_csv(StringIO(data))
Out[7]: 
  col1 col2  col3
0    a    b     1
1    a    b     2
2    c    d     3

In [8]: pd.read_csv(StringIO(data), skiprows=lambda x: x % 2 != 0)
Out[8]: 
  col1 col2  col3
0    a    b     2
skipfooterint, default 0

Number of lines at bottom of file to skip (unsupported with engine=’c’).

nrowsint, default None

Number of rows of file to read. Useful for reading pieces of large files.

low_memoryboolean, default True

Internally process the file in chunks, resulting in lower memory use while parsing, but possibly mixed type inference. To ensure no mixed types either set False, or specify the type with the dtype parameter. Note that the entire file is read into a single DataFrame regardless, use the chunksize or iterator parameter to return the data in chunks. (Only valid with C parser)

memory_mapboolean, default False

If a filepath is provided for filepath_or_buffer, map the file object directly onto memory and access the data directly from there. Using this option can improve performance because there is no longer any I/O overhead.

NA and missing data handling#

na_valuesscalar, str, list-like, or dict, default None

Additional strings to recognize as NA/NaN. If dict passed, specific per-column NA values. See na values const below for a list of the values interpreted as NaN by default.

keep_default_naboolean, default True

Whether or not to include the default NaN values when parsing the data. Depending on whether na_values is passed in, the behavior is as follows:

  • If keep_default_na is True, and na_values are specified, na_values is appended to the default NaN values used for parsing.

  • If keep_default_na is True, and na_values are not specified, only the default NaN values are used for parsing.

  • If keep_default_na is False, and na_values are specified, only the NaN values specified na_values are used for parsing.

  • If keep_default_na is False, and na_values are not specified, no strings will be parsed as NaN.

Note that if na_filter is passed in as False, the keep_default_na and na_values parameters will be ignored.

na_filterboolean, default True

Detect missing value markers (empty strings and the value of na_values). In data without any NAs, passing na_filter=False can improve the performance of reading a large file.

verboseboolean, default False

Indicate number of NA values placed in non-numeric columns.

skip_blank_linesboolean, default True

If True, skip over blank lines rather than interpreting as NaN values.

Datetime handling#

parse_datesboolean or list of ints or names or list of lists or dict, default False.
  • If True -> try parsing the index.

  • If [1, 2, 3] -> try parsing columns 1, 2, 3 each as a separate date column.

  • If [[1, 3]] -> combine columns 1 and 3 and parse as a single date column.

  • If {'foo': [1, 3]} -> parse columns 1, 3 as date and call result ‘foo’.

Note

A fast-path exists for iso8601-formatted dates.

infer_datetime_formatboolean, default False

If True and parse_dates is enabled for a column, attempt to infer the datetime format to speed up the processing.

Deprecated since version 2.0.0: A strict version of this argument is now the default, passing it has no effect.

keep_date_colboolean, default False

If True and parse_dates specifies combining multiple columns then keep the original columns.

date_parserfunction, default None

Function to use for converting a sequence of string columns to an array of datetime instances. The default uses dateutil.parser.parser to do the conversion. pandas will try to call date_parser in three different ways, advancing to the next if an exception occurs: 1) Pass one or more arrays (as defined by parse_dates) as arguments; 2) concatenate (row-wise) the string values from the columns defined by parse_dates into a single array and pass that; and 3) call date_parser once for each row using one or more strings (corresponding to the columns defined by parse_dates) as arguments.

Deprecated since version 2.0.0: Use date_format instead, or read in as object and then apply to_datetime() as-needed.

date_formatstr or dict of column -> format, default None

If used in conjunction with parse_dates, will parse dates according to this format. For anything more complex, please read in as object and then apply to_datetime() as-needed.

New in version 2.0.0.

dayfirstboolean, default False

DD/MM format dates, international and European format.

cache_datesboolean, default True

If True, use a cache of unique, converted dates to apply the datetime conversion. May produce significant speed-up when parsing duplicate date strings, especially ones with timezone offsets.

Iteration#

iteratorboolean, default False

Return TextFileReader object for iteration or getting chunks with get_chunk().

chunksizeint, default None

Return TextFileReader object for iteration. See iterating and chunking below.

Quoting, compression, and file format#

compression{'infer', 'gzip', 'bz2', 'zip', 'xz', 'zstd', None, dict}, default 'infer'

For on-the-fly decompression of on-disk data. If ‘infer’, then use gzip, bz2, zip, xz, or zstandard if filepath_or_buffer is path-like ending in ‘.gz’, ‘.bz2’, ‘.zip’, ‘.xz’, ‘.zst’, respectively, and no decompression otherwise. If using ‘zip’, the ZIP file must contain only one data file to be read in. Set to None for no decompression. Can also be a dict with key 'method' set to one of {'zip', 'gzip', 'bz2', 'zstd'} and other key-value pairs are forwarded to zipfile.ZipFile, gzip.GzipFile, bz2.BZ2File, or zstandard.ZstdDecompressor. As an example, the following could be passed for faster compression and to create a reproducible gzip archive: compression={'method': 'gzip', 'compresslevel': 1, 'mtime': 1}.

Changed in version 1.2.0: Previous versions forwarded dict entries for ‘gzip’ to gzip.open.

thousandsstr, default None

Thousands separator.

decimalstr, default '.'

Character to recognize as decimal point. E.g. use ',' for European data.

float_precisionstring, default None

Specifies which converter the C engine should use for floating-point values. The options are None for the ordinary converter, high for the high-precision converter, and round_trip for the round-trip converter.

lineterminatorstr (length 1), default None

Character to break file into lines. Only valid with C parser.

quotecharstr (length 1)

The character used to denote the start and end of a quoted item. Quoted items can include the delimiter and it will be ignored.

quotingint or csv.QUOTE_* instance, default 0

Control field quoting behavior per csv.QUOTE_* constants. Use one of QUOTE_MINIMAL (0), QUOTE_ALL (1), QUOTE_NONNUMERIC (2) or QUOTE_NONE (3).

doublequoteboolean, default True

When quotechar is specified and quoting is not QUOTE_NONE, indicate whether or not to interpret two consecutive quotechar elements inside a field as a single quotechar element.

escapecharstr (length 1), default None

One-character string used to escape delimiter when quoting is QUOTE_NONE.

commentstr, default None

Indicates remainder of line should not be parsed. If found at the beginning of a line, the line will be ignored altogether. This parameter must be a single character. Like empty lines (as long as skip_blank_lines=True), fully commented lines are ignored by the parameter header but not by skiprows. For example, if comment='#', parsing ‘#empty\na,b,c\n1,2,3’ with header=0 will result in ‘a,b,c’ being treated as the header.

encodingstr, default None

Encoding to use for UTF when reading/writing (e.g. 'utf-8'). List of Python standard encodings.

dialectstr or csv.Dialect instance, default None

If provided, this parameter will override values (default or not) for the following parameters: delimiter, doublequote, escapechar, skipinitialspace, quotechar, and quoting. If it is necessary to override values, a ParserWarning will be issued. See csv.Dialect documentation for more details.

Error handling#

on_bad_lines(‘error’, ‘warn’, ‘skip’), default ‘error’

Specifies what to do upon encountering a bad line (a line with too many fields). Allowed values are :

  • ‘error’, raise an ParserError when a bad line is encountered.

  • ‘warn’, print a warning when a bad line is encountered and skip that line.

  • ‘skip’, skip bad lines without raising or warning when they are encountered.

New in version 1.3.0.

Specifying column data types#

You can indicate the data type for the whole DataFrame or individual columns:

In [9]: import numpy as np

In [10]: data = "a,b,c,d\n1,2,3,4\n5,6,7,8\n9,10,11"

In [11]: print(data)
a,b,c,d
1,2,3,4
5,6,7,8
9,10,11

In [12]: df = pd.read_csv(StringIO(data), dtype=object)

In [13]: df
Out[13]: 
   a   b   c    d
0  1   2   3    4
1  5   6   7    8
2  9  10  11  NaN

In [14]: df["a"][0]
Out[14]: '1'

In [15]: df = pd.read_csv(StringIO(data), dtype={"b": object, "c": np.float64, "d": "Int64"})

In [16]: df.dtypes
Out[16]: 
a      int64
b     object
c    float64
d      Int64
dtype: object

Fortunately, pandas offers more than one way to ensure that your column(s) contain only one dtype. If you’re unfamiliar with these concepts, you can see here to learn more about dtypes, and here to learn more about object conversion in pandas.

For instance, you can use the converters argument of read_csv():

In [17]: data = "col_1\n1\n2\n'A'\n4.22"

In [18]: df = pd.read_csv(StringIO(data), converters={"col_1": str})

In [19]: df
Out[19]: 
  col_1
0     1
1     2
2   'A'
3  4.22

In [20]: df["col_1"].apply(type).value_counts()
Out[20]: 
col_1
<class 'str'>    4
Name: count, dtype: int64

Or you can use the to_numeric() function to coerce the dtypes after reading in the data,

In [21]: df2 = pd.read_csv(StringIO(data))

In [22]: df2["col_1"] = pd.to_numeric(df2["col_1"], errors="coerce")

In [23]: df2
Out[23]: 
   col_1
0   1.00
1   2.00
2    NaN
3   4.22

In [24]: df2["col_1"].apply(type).value_counts()
Out[24]: 
col_1
<class 'float'>    4
Name: count, dtype: int64

which will convert all valid parsing to floats, leaving the invalid parsing as NaN.

Ultimately, how you deal with reading in columns containing mixed dtypes depends on your specific needs. In the case above, if you wanted to NaN out the data anomalies, then to_numeric() is probably your best option. However, if you wanted for all the data to be coerced, no matter the type, then using the converters argument of read_csv() would certainly be worth trying.

Note

In some cases, reading in abnormal data with columns containing mixed dtypes will result in an inconsistent dataset. If you rely on pandas to infer the dtypes of your columns, the parsing engine will go and infer the dtypes for different chunks of the data, rather than the whole dataset at once. Consequently, you can end up with column(s) with mixed dtypes. For example,

In [25]: col_1 = list(range(500000)) + ["a", "b"] + list(range(500000))

In [26]: df = pd.DataFrame({"col_1": col_1})

In [27]: df.to_csv("foo.csv")

In [28]: mixed_df = pd.read_csv("foo.csv")

In [29]: mixed_df["col_1"].apply(type).value_counts()
Out[29]: 
col_1
<class 'int'>    737858
<class 'str'>    262144
Name: count, dtype: int64

In [30]: mixed_df["col_1"].dtype
Out[30]: dtype('O')

will result with mixed_df containing an int dtype for certain chunks of the column, and str for others due to the mixed dtypes from the data that was read in. It is important to note that the overall column will be marked with a dtype of object, which is used for columns with mixed dtypes.

Setting dtype_backend="numpy_nullable" will result in nullable dtypes for every column.

In [31]: data = """a,b,c,d,e,f,g,h,i,j
   ....: 1,2.5,True,a,,,,,12-31-2019,
   ....: 3,4.5,False,b,6,7.5,True,a,12-31-2019,
   ....: """
   ....: 

In [32]: df = pd.read_csv(StringIO(data), dtype_backend="numpy_nullable", parse_dates=["i"])

In [33]: df
Out[33]: 
   a    b      c  d     e     f     g     h          i     j
0  1  2.5   True  a  <NA>  <NA>  <NA>  <NA> 2019-12-31  <NA>
1  3  4.5  False  b     6   7.5  True     a 2019-12-31  <NA>

In [34]: df.dtypes
Out[34]: 
a             Int64
b           Float64
c           boolean
d    string[python]
e             Int64
f           Float64
g           boolean
h    string[python]
i    datetime64[ns]
j             Int64
dtype: object

Specifying categorical dtype#

Categorical columns can be parsed directly by specifying dtype='category' or dtype=CategoricalDtype(categories, ordered).

In [35]: data = "col1,col2,col3\na,b,1\na,b,2\nc,d,3"

In [36]: pd.read_csv(StringIO(data))
Out[36]: 
  col1 col2  col3
0    a    b     1
1    a    b     2
2    c    d     3

In [37]: pd.read_csv(StringIO(data)).dtypes
Out[37]: 
col1    object
col2    object
col3     int64
dtype: object

In [38]: pd.read_csv(StringIO(data), dtype="category").dtypes
Out[38]: 
col1    category
col2    category
col3    category
dtype: object

Individual columns can be parsed as a Categorical using a dict specification:

In [39]: pd.read_csv(StringIO(data), dtype={"col1": "category"}).dtypes
Out[39]: 
col1    category
col2      object
col3       int64
dtype: object

Specifying dtype='category' will result in an unordered Categorical whose categories are the unique values observed in the data. For more control on the categories and order, create a CategoricalDtype ahead of time, and pass that for that column’s dtype.

In [40]: from pandas.api.types import CategoricalDtype

In [41]: dtype = CategoricalDtype(["d", "c", "b", "a"], ordered=True)

In [42]: pd.read_csv(StringIO(data), dtype={"col1": dtype}).dtypes
Out[42]: 
col1    category
col2      object
col3       int64
dtype: object

When using dtype=CategoricalDtype, “unexpected” values outside of dtype.categories are treated as missing values.

In [43]: dtype = CategoricalDtype(["a", "b", "d"])  # No 'c'

In [44]: pd.read_csv(StringIO(data), dtype={"col1": dtype}).col1
Out[44]: 
0      a
1      a
2    NaN
Name: col1, dtype: category
Categories (3, object): ['a', 'b', 'd']

This matches the behavior of Categorical.set_categories().

Note

With dtype='category', the resulting categories will always be parsed as strings (object dtype). If the categories are numeric they can be converted using the to_numeric() function, or as appropriate, another converter such as to_datetime().

When dtype is a CategoricalDtype with homogeneous categories ( all numeric, all datetimes, etc.), the conversion is done automatically.

In [45]: df = pd.read_csv(StringIO(data), dtype="category")

In [46]: df.dtypes
Out[46]: 
col1    category
col2    category
col3    category
dtype: object

In [47]: df["col3"]
Out[47]: 
0    1
1    2
2    3
Name: col3, dtype: category
Categories (3, object): ['1', '2', '3']

In [48]: new_categories = pd.to_numeric(df["col3"].cat.categories)

In [49]: df["col3"] = df["col3"].cat.rename_categories(new_categories)

In [50]: df["col3"]
Out[50]: 
0    1
1    2
2    3
Name: col3, dtype: category
Categories (3, int64): [1, 2, 3]

Naming and using columns#

Handling column names#

A file may or may not have a header row. pandas assumes the first row should be used as the column names:

In [51]: data = "a,b,c\n1,2,3\n4,5,6\n7,8,9"

In [52]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9

In [53]: pd.read_csv(StringIO(data))
Out[53]: 
   a  b  c
0  1  2  3
1  4  5  6
2  7  8  9

By specifying the names argument in conjunction with header you can indicate other names to use and whether or not to throw away the header row (if any):

In [54]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9

In [55]: pd.read_csv(StringIO(data), names=["foo", "bar", "baz"], header=0)
Out[55]: 
   foo  bar  baz
0    1    2    3
1    4    5    6
2    7    8    9

In [56]: pd.read_csv(StringIO(data), names=["foo", "bar", "baz"], header=None)
Out[56]: 
  foo bar baz
0   a   b   c
1   1   2   3
2   4   5   6
3   7   8   9

If the header is in a row other than the first, pass the row number to header. This will skip the preceding rows:

In [57]: data = "skip this skip it\na,b,c\n1,2,3\n4,5,6\n7,8,9"

In [58]: pd.read_csv(StringIO(data), header=1)
Out[58]: 
   a  b  c
0  1  2  3
1  4  5  6
2  7  8  9

Note

Default behavior is to infer the column names: if no names are passed the behavior is identical to header=0 and column names are inferred from the first non-blank line of the file, if column names are passed explicitly then the behavior is identical to header=None.

Duplicate names parsing#

If the file or header contains duplicate names, pandas will by default distinguish between them so as to prevent overwriting data:

In [59]: data = "a,b,a\n0,1,2\n3,4,5"

In [60]: pd.read_csv(StringIO(data))
Out[60]: 
   a  b  a.1
0  0  1    2
1  3  4    5

There is no more duplicate data because duplicate columns ‘X’, …, ‘X’ become ‘X’, ‘X.1’, …, ‘X.N’.

Filtering columns (usecols)#

The usecols argument allows you to select any subset of the columns in a file, either using the column names, position numbers or a callable:

In [61]: data = "a,b,c,d\n1,2,3,foo\n4,5,6,bar\n7,8,9,baz"

In [62]: pd.read_csv(StringIO(data))
Out[62]: 
   a  b  c    d
0  1  2  3  foo
1  4  5  6  bar
2  7  8  9  baz

In [63]: pd.read_csv(StringIO(data), usecols=["b", "d"])
Out[63]: 
   b    d
0  2  foo
1  5  bar
2  8  baz

In [64]: pd.read_csv(StringIO(data), usecols=[0, 2, 3])
Out[64]: 
   a  c    d
0  1  3  foo
1  4  6  bar
2  7  9  baz

In [65]: pd.read_csv(StringIO(data), usecols=lambda x: x.upper() in ["A", "C"])
Out[65]: 
   a  c
0  1  3
1  4  6
2  7  9

The usecols argument can also be used to specify which columns not to use in the final result:

In [66]: pd.read_csv(StringIO(data), usecols=lambda x: x not in ["a", "c"])
Out[66]: 
   b    d
0  2  foo
1  5  bar
2  8  baz

In this case, the callable is specifying that we exclude the “a” and “c” columns from the output.

Comments and empty lines#

Ignoring line comments and empty lines#

If the comment parameter is specified, then completely commented lines will be ignored. By default, completely blank lines will be ignored as well.

In [67]: data = "\na,b,c\n  \n# commented line\n1,2,3\n\n4,5,6"

In [68]: print(data)

a,b,c
  
# commented line
1,2,3

4,5,6

In [69]: pd.read_csv(StringIO(data), comment="#")
Out[69]: 
   a  b  c
0  1  2  3
1  4  5  6

If skip_blank_lines=False, then read_csv will not ignore blank lines:

In [70]: data = "a,b,c\n\n1,2,3\n\n\n4,5,6"

In [71]: pd.read_csv(StringIO(data), skip_blank_lines=False)
Out[71]: 
     a    b    c
0  NaN  NaN  NaN
1  1.0  2.0  3.0
2  NaN  NaN  NaN
3  NaN  NaN  NaN
4  4.0  5.0  6.0

Warning

The presence of ignored lines might create ambiguities involving line numbers; the parameter header uses row numbers (ignoring commented/empty lines), while skiprows uses line numbers (including commented/empty lines):

In [72]: data = "#comment\na,b,c\nA,B,C\n1,2,3"

In [73]: pd.read_csv(StringIO(data), comment="#", header=1)
Out[73]: 
   A  B  C
0  1  2  3

In [74]: data = "A,B,C\n#comment\na,b,c\n1,2,3"

In [75]: pd.read_csv(StringIO(data), comment="#", skiprows=2)
Out[75]: 
   a  b  c
0  1  2  3

If both header and skiprows are specified, header will be relative to the end of skiprows. For example:

In [76]: data = (
   ....:     "# empty\n"
   ....:     "# second empty line\n"
   ....:     "# third emptyline\n"
   ....:     "X,Y,Z\n"
   ....:     "1,2,3\n"
   ....:     "A,B,C\n"
   ....:     "1,2.,4.\n"
   ....:     "5.,NaN,10.0\n"
   ....: )
   ....: 

In [77]: print(data)
# empty
# second empty line
# third emptyline
X,Y,Z
1,2,3
A,B,C
1,2.,4.
5.,NaN,10.0


In [78]: pd.read_csv(StringIO(data), comment="#", skiprows=4, header=1)
Out[78]: 
     A    B     C
0  1.0  2.0   4.0
1  5.0  NaN  10.0

Comments#

Sometimes comments or meta data may be included in a file:

In [79]: data = (
   ....:     "ID,level,category\n"
   ....:     "Patient1,123000,x # really unpleasant\n"
   ....:     "Patient2,23000,y # wouldn't take his medicine\n"
   ....:     "Patient3,1234018,z # awesome"
   ....: )
   ....: 

In [80]: with open("tmp.csv", "w") as fh:
   ....:     fh.write(data)
   ....: 

In [81]: print(open("tmp.csv").read())
ID,level,category
Patient1,123000,x # really unpleasant
Patient2,23000,y # wouldn't take his medicine
Patient3,1234018,z # awesome

By default, the parser includes the comments in the output:

In [82]: df = pd.read_csv("tmp.csv")

In [83]: df
Out[83]: 
         ID    level                        category
0  Patient1   123000           x # really unpleasant
1  Patient2    23000  y # wouldn't take his medicine
2  Patient3  1234018                     z # awesome

We can suppress the comments using the comment keyword:

In [84]: df = pd.read_csv("tmp.csv", comment="#")

In [85]: df
Out[85]: 
         ID    level category
0  Patient1   123000       x 
1  Patient2    23000       y 
2  Patient3  1234018       z 

Dealing with Unicode data#

The encoding argument should be used for encoded unicode data, which will result in byte strings being decoded to unicode in the result:

In [86]: from io import BytesIO

In [87]: data = b"word,length\n" b"Tr\xc3\xa4umen,7\n" b"Gr\xc3\xbc\xc3\x9fe,5"

In [88]: data = data.decode("utf8").encode("latin-1")

In [89]: df = pd.read_csv(BytesIO(data), encoding="latin-1")

In [90]: df
Out[90]: 
      word  length
0  Träumen       7
1    Grüße       5

In [91]: df["word"][1]
Out[91]: 'Grüße'

Some formats which encode all characters as multiple bytes, like UTF-16, won’t parse correctly at all without specifying the encoding. Full list of Python standard encodings.

Index columns and trailing delimiters#

If a file has one more column of data than the number of column names, the first column will be used as the DataFrame’s row names:

In [92]: data = "a,b,c\n4,apple,bat,5.7\n8,orange,cow,10"

In [93]: pd.read_csv(StringIO(data))
Out[93]: 
        a    b     c
4   apple  bat   5.7
8  orange  cow  10.0
In [94]: data = "index,a,b,c\n4,apple,bat,5.7\n8,orange,cow,10"

In [95]: pd.read_csv(StringIO(data), index_col=0)
Out[95]: 
            a    b     c
index                   
4       apple  bat   5.7
8      orange  cow  10.0

Ordinarily, you can achieve this behavior using the index_col option.

There are some exception cases when a file has been prepared with delimiters at the end of each data line, confusing the parser. To explicitly disable the index column inference and discard the last column, pass index_col=False:

In [96]: data = "a,b,c\n4,apple,bat,\n8,orange,cow,"

In [97]: print(data)
a,b,c
4,apple,bat,
8,orange,cow,

In [98]: pd.read_csv(StringIO(data))
Out[98]: 
        a    b   c
4   apple  bat NaN
8  orange  cow NaN

In [99]: pd.read_csv(StringIO(data), index_col=False)
Out[99]: 
   a       b    c
0  4   apple  bat
1  8  orange  cow

If a subset of data is being parsed using the usecols option, the index_col specification is based on that subset, not the original data.

In [100]: data = "a,b,c\n4,apple,bat,\n8,orange,cow,"

In [101]: print(data)
a,b,c
4,apple,bat,
8,orange,cow,

In [102]: pd.read_csv(StringIO(data), usecols=["b", "c"])
Out[102]: 
     b   c
4  bat NaN
8  cow NaN

In [103]: pd.read_csv(StringIO(data), usecols=["b", "c"], index_col=0)
Out[103]: 
     b   c
4  bat NaN
8  cow NaN

Date Handling#

Specifying date columns#

To better facilitate working with datetime data, read_csv() uses the keyword arguments parse_dates and date_format to allow users to specify a variety of columns and date/time formats to turn the input text data into datetime objects.

The simplest case is to just pass in parse_dates=True:

In [104]: with open("foo.csv", mode="w") as f:
   .....:     f.write("date,A,B,C\n20090101,a,1,2\n20090102,b,3,4\n20090103,c,4,5")
   .....: 

# Use a column as an index, and parse it as dates.
In [105]: df = pd.read_csv("foo.csv", index_col=0, parse_dates=True)

In [106]: df
Out[106]: 
            A  B  C
date               
2009-01-01  a  1  2
2009-01-02  b  3  4
2009-01-03  c  4  5

# These are Python datetime objects
In [107]: df.index
Out[107]: DatetimeIndex(['2009-01-01', '2009-01-02', '2009-01-03'], dtype='datetime64[ns]', name='date', freq=None)

It is often the case that we may want to store date and time data separately, or store various date fields separately. the parse_dates keyword can be used to specify a combination of columns to parse the dates and/or times from.

You can specify a list of column lists to parse_dates, the resulting date columns will be prepended to the output (so as to not affect the existing column order) and the new column names will be the concatenation of the component column names:

In [108]: data = (
   .....:     "KORD,19990127, 19:00:00, 18:56:00, 0.8100\n"
   .....:     "KORD,19990127, 20:00:00, 19:56:00, 0.0100\n"
   .....:     "KORD,19990127, 21:00:00, 20:56:00, -0.5900\n"
   .....:     "KORD,19990127, 21:00:00, 21:18:00, -0.9900\n"
   .....:     "KORD,19990127, 22:00:00, 21:56:00, -0.5900\n"
   .....:     "KORD,19990127, 23:00:00, 22:56:00, -0.5900"
   .....: )
   .....: 

In [109]: with open("tmp.csv", "w") as fh:
   .....:     fh.write(data)
   .....: 

In [110]: df = pd.read_csv("tmp.csv", header=None, parse_dates=[[1, 2], [1, 3]])

In [111]: df
Out[111]: 
                  1_2                 1_3     0     4
0 1999-01-27 19:00:00 1999-01-27 18:56:00  KORD  0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00  KORD  0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00  KORD -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00  KORD -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00  KORD -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00  KORD -0.59

By default the parser removes the component date columns, but you can choose to retain them via the keep_date_col keyword:

In [112]: df = pd.read_csv(
   .....:     "tmp.csv", header=None, parse_dates=[[1, 2], [1, 3]], keep_date_col=True
   .....: )
   .....: 

In [113]: df
Out[113]: 
                  1_2                 1_3     0  ...          2          3     4
0 1999-01-27 19:00:00 1999-01-27 18:56:00  KORD  ...   19:00:00   18:56:00  0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00  KORD  ...   20:00:00   19:56:00  0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00  KORD  ...   21:00:00   20:56:00 -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00  KORD  ...   21:00:00   21:18:00 -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00  KORD  ...   22:00:00   21:56:00 -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00  KORD  ...   23:00:00   22:56:00 -0.59

[6 rows x 7 columns]

Note that if you wish to combine multiple columns into a single date column, a nested list must be used. In other words, parse_dates=[1, 2] indicates that the second and third columns should each be parsed as separate date columns while parse_dates=[[1, 2]] means the two columns should be parsed into a single column.

You can also use a dict to specify custom name columns:

In [114]: date_spec = {"nominal": [1, 2], "actual": [1, 3]}

In [115]: df = pd.read_csv("tmp.csv", header=None, parse_dates=date_spec)

In [116]: df
Out[116]: 
              nominal              actual     0     4
0 1999-01-27 19:00:00 1999-01-27 18:56:00  KORD  0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00  KORD  0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00  KORD -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00  KORD -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00  KORD -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00  KORD -0.59

It is important to remember that if multiple text columns are to be parsed into a single date column, then a new column is prepended to the data. The index_col specification is based off of this new set of columns rather than the original data columns:

In [117]: date_spec = {"nominal": [1, 2], "actual": [1, 3]}

In [118]: df = pd.read_csv(
   .....:     "tmp.csv", header=None, parse_dates=date_spec, index_col=0
   .....: )  # index is the nominal column
   .....: 

In [119]: df
Out[119]: 
                                 actual     0     4
nominal                                            
1999-01-27 19:00:00 1999-01-27 18:56:00  KORD  0.81
1999-01-27 20:00:00 1999-01-27 19:56:00  KORD  0.01
1999-01-27 21:00:00 1999-01-27 20:56:00  KORD -0.59
1999-01-27 21:00:00 1999-01-27 21:18:00  KORD -0.99
1999-01-27 22:00:00 1999-01-27 21:56:00  KORD -0.59
1999-01-27 23:00:00 1999-01-27 22:56:00  KORD -0.59

Note

If a column or index contains an unparsable date, the entire column or index will be returned unaltered as an object data type. For non-standard datetime parsing, use to_datetime() after pd.read_csv.

Note

read_csv has a fast_path for parsing datetime strings in iso8601 format, e.g “2000-01-01T00:01:02+00:00” and similar variations. If you can arrange for your data to store datetimes in this format, load times will be significantly faster, ~20x has been observed.

Date parsing functions#

Finally, the parser allows you to specify a custom date_format. Performance-wise, you should try these methods of parsing dates in order:

  1. If you know the format, use date_format, e.g.: date_format="%d/%m/%Y" or date_format={column_name: "%d/%m/%Y"}.

  2. If you different formats for different columns, or want to pass any extra options (such as utc) to to_datetime, then you should read in your data as object dtype, and then use to_datetime.

Parsing a CSV with mixed timezones#

pandas cannot natively represent a column or index with mixed timezones. If your CSV file contains columns with a mixture of timezones, the default result will be an object-dtype column with strings, even with parse_dates. To parse the mixed-timezone values as a datetime column, read in as object dtype and then call to_datetime() with utc=True.

In [120]: content = """\
   .....: a
   .....: 2000-01-01T00:00:00+05:00
   .....: 2000-01-01T00:00:00+06:00"""
   .....: 

In [121]: df = pd.read_csv(StringIO(content))

In [122]: df["a"] = pd.to_datetime(df["a"], utc=True)

In [123]: df["a"]
Out[123]: 
0   1999-12-31 19:00:00+00:00
1   1999-12-31 18:00:00+00:00
Name: a, dtype: datetime64[ns, UTC]

Inferring datetime format#

Here are some examples of datetime strings that can be guessed (all representing December 30th, 2011 at 00:00:00):

  • “20111230”

  • “2011/12/30”

  • “20111230 00:00:00”

  • “12/30/2011 00:00:00”

  • “30/Dec/2011 00:00:00”

  • “30/December/2011 00:00:00”

Note that format inference is sensitive to dayfirst. With dayfirst=True, it will guess “01/12/2011” to be December 1st. With dayfirst=False (default) it will guess “01/12/2011” to be January 12th.

If you try to parse a column of date strings, pandas will attempt to guess the format from the first non-NaN element, and will then parse the rest of the column with that format. If pandas fails to guess the format (for example if your first string is '01 December US/Pacific 2000'), then a warning will be raised and each row will be parsed individually by dateutil.parser.parse. The safest way to parse dates is to explicitly set format=.

In [124]: df = pd.read_csv(
   .....:     "foo.csv",
   .....:     index_col=0,
   .....:     parse_dates=True,
   .....: )
   .....: 

In [125]: df
Out[125]: 
            A  B  C
date               
2009-01-01  a  1  2
2009-01-02  b  3  4
2009-01-03  c  4  5

In the case that you have mixed datetime formats within the same column, you can pass format='mixed'

In [126]: data = StringIO("date\n12 Jan 2000\n2000-01-13\n")

In [127]: df = pd.read_csv(data)

In [128]: df['date'] = pd.to_datetime(df['date'], format='mixed')

In [129]: df
Out[129]: 
        date
0 2000-01-12
1 2000-01-13

or, if your datetime formats are all ISO8601 (possibly not identically-formatted):

In [130]: data = StringIO("date\n2020-01-01\n2020-01-01 03:00\n")

In [131]: df = pd.read_csv(data)

In [132]: df['date'] = pd.to_datetime(df['date'], format='ISO8601')

In [133]: df
Out[133]: 
                 date
0 2020-01-01 00:00:00
1 2020-01-01 03:00:00

International date formats#

While US date formats tend to be MM/DD/YYYY, many international formats use DD/MM/YYYY instead. For convenience, a dayfirst keyword is provided:

In [134]: data = "date,value,cat\n1/6/2000,5,a\n2/6/2000,10,b\n3/6/2000,15,c"

In [135]: print(data)
date,value,cat
1/6/2000,5,a
2/6/2000,10,b
3/6/2000,15,c

In [136]: with open("tmp.csv", "w") as fh:
   .....:     fh.write(data)
   .....: 

In [137]: pd.read_csv("tmp.csv", parse_dates=[0])
Out[137]: 
        date  value cat
0 2000-01-06      5   a
1 2000-02-06     10   b
2 2000-03-06     15   c

In [138]: pd.read_csv("tmp.csv", dayfirst=True, parse_dates=[0])
Out[138]: 
        date  value cat
0 2000-06-01      5   a
1 2000-06-02     10   b
2 2000-06-03     15   c

Writing CSVs to binary file objects#

New in version 1.2.0.

df.to_csv(..., mode="wb") allows writing a CSV to a file object opened binary mode. In most cases, it is not necessary to specify mode as Pandas will auto-detect whether the file object is opened in text or binary mode.

In [139]: import io

In [140]: data = pd.DataFrame([0, 1, 2])

In [141]: buffer = io.BytesIO()

In [142]: data.to_csv(buffer, encoding="utf-8", compression="gzip")

Specifying method for floating-point conversion#

The parameter float_precision can be specified in order to use a specific floating-point converter during parsing with the C engine. The options are the ordinary converter, the high-precision converter, and the round-trip converter (which is guaranteed to round-trip values after writing to a file). For example:

In [143]: val = "0.3066101993807095471566981359501369297504425048828125"

In [144]: data = "a,b,c\n1,2,{0}".format(val)

In [145]: abs(
   .....:     pd.read_csv(
   .....:         StringIO(data),
   .....:         engine="c",
   .....:         float_precision=None,
   .....:     )["c"][0] - float(val)
   .....: )
   .....: 
Out[145]: 5.551115123125783e-17

In [146]: abs(
   .....:     pd.read_csv(
   .....:         StringIO(data),
   .....:         engine="c",
   .....:         float_precision="high",
   .....:     )["c"][0] - float(val)
   .....: )
   .....: 
Out[146]: 5.551115123125783e-17

In [147]: abs(
   .....:     pd.read_csv(StringIO(data), engine="c", float_precision="round_trip")["c"][0]
   .....:     - float(val)
   .....: )
   .....: 
Out[147]: 0.0

Thousand separators#

For large numbers that have been written with a thousands separator, you can set the thousands keyword to a string of length 1 so that integers will be parsed correctly:

By default, numbers with a thousands separator will be parsed as strings:

In [148]: data = (
   .....:     "ID|level|category\n"
   .....:     "Patient1|123,000|x\n"
   .....:     "Patient2|23,000|y\n"
   .....:     "Patient3|1,234,018|z"
   .....: )
   .....: 

In [149]: with open("tmp.csv", "w") as fh:
   .....:     fh.write(data)
   .....: 

In [150]: df = pd.read_csv("tmp.csv", sep="|")

In [151]: df
Out[151]: 
         ID      level category
0  Patient1    123,000        x
1  Patient2     23,000        y
2  Patient3  1,234,018        z

In [152]: df.level.dtype
Out[152]: dtype('O')

The thousands keyword allows integers to be parsed correctly:

In [153]: df = pd.read_csv("tmp.csv", sep="|", thousands=",")

In [154]: df
Out[154]: 
         ID    level category
0  Patient1   123000        x
1  Patient2    23000        y
2  Patient3  1234018        z

In [155]: df.level.dtype
Out[155]: dtype('int64')

NA values#

To control which values are parsed as missing values (which are signified by NaN), specify a string in na_values. If you specify a list of strings, then all values in it are considered to be missing values. If you specify a number (a float, like 5.0 or an integer like 5), the corresponding equivalent values will also imply a missing value (in this case effectively [5.0, 5] are recognized as NaN).

To completely override the default values that are recognized as missing, specify keep_default_na=False.

The default NaN recognized values are ['-1.#IND', '1.#QNAN', '1.#IND', '-1.#QNAN', '#N/A N/A', '#N/A', 'N/A', 'n/a', 'NA', '<NA>', '#NA', 'NULL', 'null', 'NaN', '-NaN', 'nan', '-nan', 'None', ''].

Let us consider some examples:

pd.read_csv("path_to_file.csv", na_values=[5])

In the example above 5 and 5.0 will be recognized as NaN, in addition to the defaults. A string will first be interpreted as a numerical 5, then as a NaN.

pd.read_csv("path_to_file.csv", keep_default_na=False, na_values=[""])

Above, only an empty field will be recognized as NaN.

pd.read_csv("path_to_file.csv", keep_default_na=False, na_values=["NA", "0"])

Above, both NA and 0 as strings are NaN.

pd.read_csv("path_to_file.csv", na_values=["Nope"])

The default values, in addition to the string "Nope" are recognized as NaN.

Infinity#

inf like values will be parsed as np.inf (positive infinity), and -inf as -np.inf (negative infinity). These will ignore the case of the value, meaning Inf, will also be parsed as np.inf.

Boolean values#

The common values True, False, TRUE, and FALSE are all recognized as boolean. Occasionally you might want to recognize other values as being boolean. To do this, use the true_values and false_values options as follows:

In [156]: data = "a,b,c\n1,Yes,2\n3,No,4"

In [157]: print(data)
a,b,c
1,Yes,2
3,No,4

In [158]: pd.read_csv(StringIO(data))
Out[158]: 
   a    b  c
0  1  Yes  2
1  3   No  4

In [159]: pd.read_csv(StringIO(data), true_values=["Yes"], false_values=["No"])
Out[159]: 
   a      b  c
0  1   True  2
1  3  False  4

Handling “bad” lines#

Some files may have malformed lines with too few fields or too many. Lines with too few fields will have NA values filled in the trailing fields. Lines with too many fields will raise an error by default:

In [160]: data = "a,b,c\n1,2,3\n4,5,6,7\n8,9,10"

In [161]: pd.read_csv(StringIO(data))
---------------------------------------------------------------------------
ParserError                               Traceback (most recent call last)
Cell In[161], line 1
----> 1 pd.read_csv(StringIO(data))

File ~/work/pandas/pandas/pandas/io/parsers/readers.py:948, in read_csv(filepath_or_buffer, sep, delimiter, header, names, index_col, usecols, dtype, engine, converters, true_values, false_values, skipinitialspace, skiprows, skipfooter, nrows, na_values, keep_default_na, na_filter, verbose, skip_blank_lines, parse_dates, infer_datetime_format, keep_date_col, date_parser, date_format, dayfirst, cache_dates, iterator, chunksize, compression, thousands, decimal, lineterminator, quotechar, quoting, doublequote, escapechar, comment, encoding, encoding_errors, dialect, on_bad_lines, delim_whitespace, low_memory, memory_map, float_precision, storage_options, dtype_backend)
    935 kwds_defaults = _refine_defaults_read(
    936     dialect,
    937     delimiter,
   (...)
    944     dtype_backend=dtype_backend,
    945 )
    946 kwds.update(kwds_defaults)
--> 948 return _read(filepath_or_buffer, kwds)

File ~/work/pandas/pandas/pandas/io/parsers/readers.py:617, in _read(filepath_or_buffer, kwds)
    614     return parser
    616 with parser:
--> 617     return parser.read(nrows)

File ~/work/pandas/pandas/pandas/io/parsers/readers.py:1748, in TextFileReader.read(self, nrows)
   1741 nrows = validate_integer("nrows", nrows)
   1742 try:
   1743     # error: "ParserBase" has no attribute "read"
   1744     (
   1745         index,
   1746         columns,
   1747         col_dict,
-> 1748     ) = self._engine.read(  # type: ignore[attr-defined]
   1749         nrows
   1750     )
   1751 except Exception:
   1752     self.close()

File ~/work/pandas/pandas/pandas/io/parsers/c_parser_wrapper.py:234, in CParserWrapper.read(self, nrows)
    232 try:
    233     if self.low_memory:
--> 234         chunks = self._reader.read_low_memory(nrows)
    235         # destructive to chunks
    236         data = _concatenate_chunks(chunks)

File parsers.pyx:843, in pandas._libs.parsers.TextReader.read_low_memory()

File parsers.pyx:904, in pandas._libs.parsers.TextReader._read_rows()

File parsers.pyx:879, in pandas._libs.parsers.TextReader._tokenize_rows()

File parsers.pyx:890, in pandas._libs.parsers.TextReader._check_tokenize_status()

File parsers.pyx:2058, in pandas._libs.parsers.raise_parser_error()

ParserError: Error tokenizing data. C error: Expected 3 fields in line 3, saw 4

You can elect to skip bad lines:

In [162]: data = "a,b,c\n1,2,3\n4,5,6,7\n8,9,10"

In [163]: pd.read_csv(StringIO(data), on_bad_lines="skip")
Out[163]: 
   a  b   c
0  1  2   3
1  8  9  10

New in version 1.4.0.

Or pass a callable function to handle the bad line if engine="python". The bad line will be a list of strings that was split by the sep:

In [164]: external_list = []

In [165]: def bad_lines_func(line):
   .....:     external_list.append(line)
   .....:     return line[-3:]
   .....: 

In [166]: external_list
Out[166]: []

Note

The callable function will handle only a line with too many fields. Bad lines caused by other errors will be silently skipped.

In [167]: bad_lines_func = lambda line: print(line)

In [168]: data = 'name,type\nname a,a is of type a\nname b,"b\" is of type b"'

In [169]: data
Out[169]: 'name,type\nname a,a is of type a\nname b,"b" is of type b"'

In [170]: pd.read_csv(StringIO(data), on_bad_lines=bad_lines_func, engine="python")
Out[170]: 
     name            type
0  name a  a is of type a

The line was not processed in this case, as a “bad line” here is caused by an escape character.

You can also use the usecols parameter to eliminate extraneous column data that appear in some lines but not others:

In [171]: pd.read_csv(StringIO(data), usecols=[0, 1, 2])
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[171], line 1
----> 1 pd.read_csv(StringIO(data), usecols=[0, 1, 2])

File ~/work/pandas/pandas/pandas/io/parsers/readers.py:948, in read_csv(filepath_or_buffer, sep, delimiter, header, names, index_col, usecols, dtype, engine, converters, true_values, false_values, skipinitialspace, skiprows, skipfooter, nrows, na_values, keep_default_na, na_filter, verbose, skip_blank_lines, parse_dates, infer_datetime_format, keep_date_col, date_parser, date_format, dayfirst, cache_dates, iterator, chunksize, compression, thousands, decimal, lineterminator, quotechar, quoting, doublequote, escapechar, comment, encoding, encoding_errors, dialect, on_bad_lines, delim_whitespace, low_memory, memory_map, float_precision, storage_options, dtype_backend)
    935 kwds_defaults = _refine_defaults_read(
    936     dialect,
    937     delimiter,
   (...)
    944     dtype_backend=dtype_backend,
    945 )
    946 kwds.update(kwds_defaults)
--> 948 return _read(filepath_or_buffer, kwds)

File ~/work/pandas/pandas/pandas/io/parsers/readers.py:611, in _read(filepath_or_buffer, kwds)
    608 _validate_names(kwds.get("names", None))
    610 # Create the parser.
--> 611 parser = TextFileReader(filepath_or_buffer, **kwds)
    613 if chunksize or iterator:
    614     return parser

File ~/work/pandas/pandas/pandas/io/parsers/readers.py:1448, in TextFileReader.__init__(self, f, engine, **kwds)
   1445     self.options["has_index_names"] = kwds["has_index_names"]
   1447 self.handles: IOHandles | None = None
-> 1448 self._engine = self._make_engine(f, self.engine)

File ~/work/pandas/pandas/pandas/io/parsers/readers.py:1723, in TextFileReader._make_engine(self, f, engine)
   1720     raise ValueError(msg)
   1722 try:
-> 1723     return mapping[engine](f, **self.options)
   1724 except Exception:
   1725     if self.handles is not None:

File ~/work/pandas/pandas/pandas/io/parsers/c_parser_wrapper.py:155, in CParserWrapper.__init__(self, src, **kwds)
    152     # error: Cannot determine type of 'names'
    153     if len(self.names) < len(usecols):  # type: ignore[has-type]
    154         # error: Cannot determine type of 'names'
--> 155         self._validate_usecols_names(
    156             usecols,
    157             self.names,  # type: ignore[has-type]
    158         )
    160 # error: Cannot determine type of 'names'
    161 self._validate_parse_dates_presence(self.names)  # type: ignore[has-type]

File ~/work/pandas/pandas/pandas/io/parsers/base_parser.py:969, in ParserBase._validate_usecols_names(self, usecols, names)
    967 missing = [c for c in usecols if c not in names]
    968 if len(missing) > 0:
--> 969     raise ValueError(
    970         f"Usecols do not match columns, columns expected but not found: "
    971         f"{missing}"
    972     )
    974 return usecols

ValueError: Usecols do not match columns, columns expected but not found: [0, 1, 2]

In case you want to keep all data including the lines with too many fields, you can specify a sufficient number of names. This ensures that lines with not enough fields are filled with NaN.

In [172]: pd.read_csv(StringIO(data), names=['a', 'b', 'c', 'd'])
Out[172]: 
        a                b   c   d
0    name             type NaN NaN
1  name a   a is of type a NaN NaN
2  name b  b is of type b" NaN NaN

Dialect#

The dialect keyword gives greater flexibility in specifying the file format. By default it uses the Excel dialect but you can specify either the dialect name or a csv.Dialect instance.

Suppose you had data with unenclosed quotes:

In [173]: data = "label1,label2,label3\n" 'index1,"a,c,e\n' "index2,b,d,f"

In [174]: print(data)
label1,label2,label3
index1,"a,c,e
index2,b,d,f

By default, read_csv uses the Excel dialect and treats the double quote as the quote character, which causes it to fail when it finds a newline before it finds the closing double quote.

We can get around this using dialect:

In [175]: import csv

In [176]: dia = csv.excel()

In [177]: dia.quoting = csv.QUOTE_NONE

In [178]: pd.read_csv(StringIO(data), dialect=dia)
Out[178]: 
       label1 label2 label3
index1     "a      c      e
index2      b      d      f

All of the dialect options can be specified separately by keyword arguments:

In [179]: data = "a,b,c~1,2,3~4,5,6"

In [180]: pd.read_csv(StringIO(data), lineterminator="~")
Out[180]: 
   a  b  c
0  1  2  3
1  4  5  6

Another common dialect option is skipinitialspace, to skip any whitespace after a delimiter:

In [181]: data = "a, b, c\n1, 2, 3\n4, 5, 6"

In [182]: print(data)
a, b, c
1, 2, 3
4, 5, 6

In [183]: pd.read_csv(StringIO(data), skipinitialspace=True)
Out[183]: 
   a  b  c
0  1  2  3
1  4  5  6

The parsers make every attempt to “do the right thing” and not be fragile. Type inference is a pretty big deal. If a column can be coerced to integer dtype without altering the contents, the parser will do so. Any non-numeric columns will come through as object dtype as with the rest of pandas objects.

Quoting and Escape Characters#

Quotes (and other escape characters) in embedded fields can be handled in any number of ways. One way is to use backslashes; to properly parse this data, you should pass the escapechar option:

In [184]: data = 'a,b\n"hello, \\"Bob\\", nice to see you",5'

In [185]: print(data)
a,b
"hello, \"Bob\", nice to see you",5

In [186]: pd.read_csv(StringIO(data), escapechar="\\")
Out[186]: 
                               a  b
0  hello, "Bob", nice to see you  5

Files with fixed width columns#

While read_csv() reads delimited data, the read_fwf() function works with data files that have known and fixed column widths. The function parameters to read_fwf are largely the same as read_csv with two extra parameters, and a different usage of the delimiter parameter:

  • colspecs: A list of pairs (tuples) giving the extents of the fixed-width fields of each line as half-open intervals (i.e., [from, to[ ). String value ‘infer’ can be used to instruct the parser to try detecting the column specifications from the first 100 rows of the data. Default behavior, if not specified, is to infer.

  • widths: A list of field widths which can be used instead of ‘colspecs’ if the intervals are contiguous.

  • delimiter: Characters to consider as filler characters in the fixed-width file. Can be used to specify the filler character of the fields if it is not spaces (e.g., ‘~’).

Consider a typical fixed-width data file:

In [187]: data1 = (
   .....:     "id8141    360.242940   149.910199   11950.7\n"
   .....:     "id1594    444.953632   166.985655   11788.4\n"
   .....:     "id1849    364.136849   183.628767   11806.2\n"
   .....:     "id1230    413.836124   184.375703   11916.8\n"
   .....:     "id1948    502.953953   173.237159   12468.3"
   .....: )
   .....: 

In [188]: with open("bar.csv", "w") as f:
   .....:     f.write(data1)
   .....: 

In order to parse this file into a DataFrame, we simply need to supply the column specifications to the read_fwf function along with the file name:

# Column specifications are a list of half-intervals
In [189]: colspecs = [(0, 6), (8, 20), (21, 33), (34, 43)]

In [190]: df = pd.read_fwf("bar.csv", colspecs=colspecs, header=None, index_col=0)

In [191]: df
Out[191]: 
                 1           2        3
0                                      
id8141  360.242940  149.910199  11950.7
id1594  444.953632  166.985655  11788.4
id1849  364.136849  183.628767  11806.2
id1230  413.836124  184.375703  11916.8
id1948  502.953953  173.237159  12468.3

Note how the parser automatically picks column names X.<column number> when header=None argument is specified. Alternatively, you can supply just the column widths for contiguous columns:

# Widths are a list of integers
In [192]: widths = [6, 14, 13, 10]

In [193]: df = pd.read_fwf("bar.csv", widths=widths, header=None)

In [194]: df
Out[194]: 
        0           1           2        3
0  id8141  360.242940  149.910199  11950.7
1  id1594  444.953632  166.985655  11788.4
2  id1849  364.136849  183.628767  11806.2
3  id1230  413.836124  184.375703  11916.8
4  id1948  502.953953  173.237159  12468.3

The parser will take care of extra white spaces around the columns so it’s ok to have extra separation between the columns in the file.

By default, read_fwf will try to infer the file’s colspecs by using the first 100 rows of the file. It can do it only in cases when the columns are aligned and correctly separated by the provided delimiter (default delimiter is whitespace).

In [195]: df = pd.read_fwf("bar.csv", header=None, index_col=0)

In [196]: df
Out[196]: 
                 1           2        3
0                                      
id8141  360.242940  149.910199  11950.7
id1594  444.953632  166.985655  11788.4
id1849  364.136849  183.628767  11806.2
id1230  413.836124  184.375703  11916.8
id1948  502.953953  173.237159  12468.3

read_fwf supports the dtype parameter for specifying the types of parsed columns to be different from the inferred type.

In [197]: pd.read_fwf("bar.csv", header=None, index_col=0).dtypes
Out[197]: 
1    float64
2    float64
3    float64
dtype: object

In [198]: pd.read_fwf("bar.csv", header=None, dtype={2: "object"}).dtypes
Out[198]: 
0     object
1    float64
2     object
3    float64
dtype: object

Indexes#

Files with an “implicit” index column#

Consider a file with one less entry in the header than the number of data column:

In [199]: data = "A,B,C\n20090101,a,1,2\n20090102,b,3,4\n20090103,c,4,5"

In [200]: print(data)
A,B,C
20090101,a,1,2
20090102,b,3,4
20090103,c,4,5

In [201]: with open("foo.csv", "w") as f:
   .....:     f.write(data)
   .....: 

In this special case, read_csv assumes that the first column is to be used as the index of the DataFrame:

In [202]: pd.read_csv("foo.csv")
Out[202]: 
          A  B  C
20090101  a  1  2
20090102  b  3  4
20090103  c  4  5

Note that the dates weren’t automatically parsed. In that case you would need to do as before:

In [203]: df = pd.read_csv("foo.csv", parse_dates=True)

In [204]: df.index
Out[204]: DatetimeIndex(['2009-01-01', '2009-01-02', '2009-01-03'], dtype='datetime64[ns]', freq=None)

Reading an index with a MultiIndex#

Suppose you have data indexed by two columns:

In [205]: data = 'year,indiv,zit,xit\n1977,"A",1.2,.6\n1977,"B",1.5,.5'

In [206]: print(data)
year,indiv,zit,xit
1977,"A",1.2,.6
1977,"B",1.5,.5

In [207]: with open("mindex_ex.csv", mode="w") as f:
   .....:     f.write(data)
   .....: 

The index_col argument to read_csv can take a list of column numbers to turn multiple columns into a MultiIndex for the index of the returned object:

In [208]: df = pd.read_csv("mindex_ex.csv", index_col=[0, 1])

In [209]: df
Out[209]: 
            zit  xit
year indiv          
1977 A      1.2  0.6
     B      1.5  0.5

In [210]: df.loc[1977]
Out[210]: 
       zit  xit
indiv          
A      1.2  0.6
B      1.5  0.5

Reading columns with a MultiIndex#

By specifying list of row locations for the header argument, you can read in a MultiIndex for the columns. Specifying non-consecutive rows will skip the intervening rows.

In [211]: from pandas._testing import makeCustomDataframe as mkdf

In [212]: df = mkdf(5, 3, r_idx_nlevels=2, c_idx_nlevels=4)

In [213]: df.to_csv("mi.csv")

In [214]: print(open("mi.csv").read())
C0,,C_l0_g0,C_l0_g1,C_l0_g2
C1,,C_l1_g0,C_l1_g1,C_l1_g2
C2,,C_l2_g0,C_l2_g1,C_l2_g2
C3,,C_l3_g0,C_l3_g1,C_l3_g2
R0,R1,,,
R_l0_g0,R_l1_g0,R0C0,R0C1,R0C2
R_l0_g1,R_l1_g1,R1C0,R1C1,R1C2
R_l0_g2,R_l1_g2,R2C0,R2C1,R2C2
R_l0_g3,R_l1_g3,R3C0,R3C1,R3C2
R_l0_g4,R_l1_g4,R4C0,R4C1,R4C2


In [215]: pd.read_csv("mi.csv", header=[0, 1, 2, 3], index_col=[0, 1])
Out[215]: 
C0              C_l0_g0 C_l0_g1 C_l0_g2
C1              C_l1_g0 C_l1_g1 C_l1_g2
C2              C_l2_g0 C_l2_g1 C_l2_g2
C3              C_l3_g0 C_l3_g1 C_l3_g2
R0      R1                             
R_l0_g0 R_l1_g0    R0C0    R0C1    R0C2
R_l0_g1 R_l1_g1    R1C0    R1C1    R1C2
R_l0_g2 R_l1_g2    R2C0    R2C1    R2C2
R_l0_g3 R_l1_g3    R3C0    R3C1    R3C2
R_l0_g4 R_l1_g4    R4C0    R4C1    R4C2

read_csv is also able to interpret a more common format of multi-columns indices.

In [216]: data = ",a,a,a,b,c,c\n,q,r,s,t,u,v\none,1,2,3,4,5,6\ntwo,7,8,9,10,11,12"

In [217]: print(data)
,a,a,a,b,c,c
,q,r,s,t,u,v
one,1,2,3,4,5,6
two,7,8,9,10,11,12

In [218]: with open("mi2.csv", "w") as fh:
   .....:     fh.write(data)
   .....: 

In [219]: pd.read_csv("mi2.csv", header=[0, 1], index_col=0)
Out[219]: 
     a         b   c    
     q  r  s   t   u   v
one  1  2  3   4   5   6
two  7  8  9  10  11  12

Note

If an index_col is not specified (e.g. you don’t have an index, or wrote it with df.to_csv(..., index=False), then any names on the columns index will be lost.

Automatically “sniffing” the delimiter#

read_csv is capable of inferring delimited (not necessarily comma-separated) files, as pandas uses the csv.Sniffer class of the csv module. For this, you have to specify sep=None.

In [220]: df = pd.DataFrame(np.random.randn(10, 4))

In [221]: df.to_csv("tmp2.csv", sep=":", index=False)

In [222]: pd.read_csv("tmp2.csv", sep=None, engine="python")
Out[222]: 
          0         1         2         3
0  0.469112 -0.282863 -1.509059 -1.135632
1  1.212112 -0.173215  0.119209 -1.044236
2 -0.861849 -2.104569 -0.494929  1.071804
3  0.721555 -0.706771 -1.039575  0.271860
4 -0.424972  0.567020  0.276232 -1.087401
5 -0.673690  0.113648 -1.478427  0.524988
6  0.404705  0.577046 -1.715002 -1.039268
7 -0.370647 -1.157892 -1.344312  0.844885
8  1.075770 -0.109050  1.643563 -1.469388
9  0.357021 -0.674600 -1.776904 -0.968914

Reading multiple files to create a single DataFrame#

It’s best to use concat() to combine multiple files. See the cookbook for an example.

Iterating through files chunk by chunk#

Suppose you wish to iterate through a (potentially very large) file lazily rather than reading the entire file into memory, such as the following:

In [223]: df = pd.DataFrame(np.random.randn(10, 4))

In [224]: df.to_csv("tmp.csv", index=False)

In [225]: table = pd.read_csv("tmp.csv")

In [226]: table
Out[226]: 
          0         1         2         3
0 -1.294524  0.413738  0.276662 -0.472035
1 -0.013960 -0.362543 -0.006154 -0.923061
2  0.895717  0.805244 -1.206412  2.565646
3  1.431256  1.340309 -1.170299 -0.226169
4  0.410835  0.813850  0.132003 -0.827317
5 -0.076467 -1.187678  1.130127 -1.436737
6 -1.413681  1.607920  1.024180  0.569605
7  0.875906 -2.211372  0.974466 -2.006747
8 -0.410001 -0.078638  0.545952 -1.219217
9 -1.226825  0.769804 -1.281247 -0.727707

By specifying a chunksize to read_csv, the return value will be an iterable object of type TextFileReader:

In [227]: with pd.read_csv("tmp.csv", chunksize=4) as reader:
   .....:     print(reader)
   .....:     for chunk in reader:
   .....:         print(chunk)
   .....: 
<pandas.io.parsers.readers.TextFileReader object at 0x7f8749599150>
          0         1         2         3
0 -1.294524  0.413738  0.276662 -0.472035
1 -0.013960 -0.362543 -0.006154 -0.923061
2  0.895717  0.805244 -1.206412  2.565646
3  1.431256  1.340309 -1.170299 -0.226169
          0         1         2         3
4  0.410835  0.813850  0.132003 -0.827317
5 -0.076467 -1.187678  1.130127 -1.436737
6 -1.413681  1.607920  1.024180  0.569605
7  0.875906 -2.211372  0.974466 -2.006747
          0         1         2         3
8 -0.410001 -0.078638  0.545952 -1.219217
9 -1.226825  0.769804 -1.281247 -0.727707

Changed in version 1.2: read_csv/json/sas return a context-manager when iterating through a file.

Specifying iterator=True will also return the TextFileReader object:

In [228]: with pd.read_csv("tmp.csv", iterator=True) as reader:
   .....:     print(reader.get_chunk(5))
   .....: 
          0         1         2         3
0 -1.294524  0.413738  0.276662 -0.472035
1 -0.013960 -0.362543 -0.006154 -0.923061
2  0.895717  0.805244 -1.206412  2.565646
3  1.431256  1.340309 -1.170299 -0.226169
4  0.410835  0.813850  0.132003 -0.827317

Specifying the parser engine#

Pandas currently supports three engines, the C engine, the python engine, and an experimental pyarrow engine (requires the pyarrow package). In general, the pyarrow engine is fastest on larger workloads and is equivalent in speed to the C engine on most other workloads. The python engine tends to be slower than the pyarrow and C engines on most workloads. However, the pyarrow engine is much less robust than the C engine, which lacks a few features compared to the Python engine.

Where possible, pandas uses the C parser (specified as engine='c'), but it may fall back to Python if C-unsupported options are specified.

Currently, options unsupported by the C and pyarrow engines include:

  • sep other than a single character (e.g. regex separators)

  • skipfooter

  • sep=None with delim_whitespace=False

Specifying any of the above options will produce a ParserWarning unless the python engine is selected explicitly using engine='python'.

Options that are unsupported by the pyarrow engine which are not covered by the list above include:

  • float_precision

  • chunksize

  • comment

  • nrows

  • thousands

  • memory_map

  • dialect

  • on_bad_lines

  • delim_whitespace

  • quoting

  • lineterminator

  • converters

  • decimal

  • iterator

  • dayfirst

  • infer_datetime_format

  • verbose

  • skipinitialspace

  • low_memory

Specifying these options with engine='pyarrow' will raise a ValueError.

Reading/writing remote files#

You can pass in a URL to read or write remote files to many of pandas’ IO functions - the following example shows reading a CSV file:

df = pd.read_csv("https://download.bls.gov/pub/time.series/cu/cu.item", sep="\t")

New in version 1.3.0.

A custom header can be sent alongside HTTP(s) requests by passing a dictionary of header key value mappings to the storage_options keyword argument as shown below:

headers = {"User-Agent": "pandas"}
df = pd.read_csv(
    "https://download.bls.gov/pub/time.series/cu/cu.item",
    sep="\t",
    storage_options=headers
)

All URLs which are not local files or HTTP(s) are handled by fsspec, if installed, and its various filesystem implementations (including Amazon S3, Google Cloud, SSH, FTP, webHDFS…). Some of these implementations will require additional packages to be installed, for example S3 URLs require the s3fs library:

df = pd.read_json("s3://pandas-test/adatafile.json")

When dealing with remote storage systems, you might need extra configuration with environment variables or config files in special locations. For example, to access data in your S3 bucket, you will need to define credentials in one of the several ways listed in the S3Fs documentation. The same is true for several of the storage backends, and you should follow the links at fsimpl1 for implementations built into fsspec and fsimpl2 for those not included in the main fsspec distribution.

You can also pass parameters directly to the backend driver. Since fsspec does not utilize the AWS_S3_HOST environment variable, we can directly define a dictionary containing the endpoint_url and pass the object into the storage option parameter:

storage_options = {"client_kwargs": {"endpoint_url": "http://127.0.0.1:5555"}}}
df = pd.read_json("s3://pandas-test/test-1", storage_options=storage_options)

More sample configurations and documentation can be found at S3Fs documentation.

If you do not have S3 credentials, you can still access public data by specifying an anonymous connection, such as

New in version 1.2.0.

pd.read_csv(
    "s3://ncei-wcsd-archive/data/processed/SH1305/18kHz/SaKe2013"
    "-D20130523-T080854_to_SaKe2013-D20130523-T085643.csv",
    storage_options={"anon": True},
)

fsspec also allows complex URLs, for accessing data in compressed archives, local caching of files, and more. To locally cache the above example, you would modify the call to

pd.read_csv(
    "simplecache::s3://ncei-wcsd-archive/data/processed/SH1305/18kHz/"
    "SaKe2013-D20130523-T080854_to_SaKe2013-D20130523-T085643.csv",
    storage_options={"s3": {"anon": True}},
)

where we specify that the “anon” parameter is meant for the “s3” part of the implementation, not to the caching implementation. Note that this caches to a temporary directory for the duration of the session only, but you can also specify a permanent store.

Writing out data#

Writing to CSV format#

The Series and DataFrame objects have an instance method to_csv which allows storing the contents of the object as a comma-separated-values file. The function takes a number of arguments. Only the first is required.

  • path_or_buf: A string path to the file to write or a file object. If a file object it must be opened with newline=''

  • sep : Field delimiter for the output file (default “,”)

  • na_rep: A string representation of a missing value (default ‘’)

  • float_format: Format string for floating point numbers

  • columns: Columns to write (default None)

  • header: Whether to write out the column names (default True)

  • index: whether to write row (index) names (default True)

  • index_label: Column label(s) for index column(s) if desired. If None (default), and header and index are True, then the index names are used. (A sequence should be given if the DataFrame uses MultiIndex).

  • mode : Python write mode, default ‘w’

  • encoding: a string representing the encoding to use if the contents are non-ASCII, for Python versions prior to 3

  • lineterminator: Character sequence denoting line end (default os.linesep)

  • quoting: Set quoting rules as in csv module (default csv.QUOTE_MINIMAL). Note that if you have set a float_format then floats are converted to strings and csv.QUOTE_NONNUMERIC will treat them as non-numeric

  • quotechar: Character used to quote fields (default ‘”’)

  • doublequote: Control quoting of quotechar in fields (default True)

  • escapechar: Character used to escape sep and quotechar when appropriate (default None)

  • chunksize: Number of rows to write at a time

  • date_format: Format string for datetime objects

Writing a formatted string#

The DataFrame object has an instance method to_string which allows control over the string representation of the object. All arguments are optional:

  • buf default None, for example a StringIO object

  • columns default None, which columns to write

  • col_space default None, minimum width of each column.

  • na_rep default NaN, representation of NA value

  • formatters default None, a dictionary (by column) of functions each of which takes a single argument and returns a formatted string

  • float_format default None, a function which takes a single (float) argument and returns a formatted string; to be applied to floats in the DataFrame.

  • sparsify default True, set to False for a DataFrame with a hierarchical index to print every MultiIndex key at each row.

  • index_names default True, will print the names of the indices

  • index default True, will print the index (ie, row labels)

  • header default True, will print the column labels

  • justify default left, will print column headers left- or right-justified

The Series object also has a to_string method, but with only the buf, na_rep, float_format arguments. There is also a length argument which, if set to True, will additionally output the length of the Series.

JSON#

Read and write JSON format files and strings.

Writing JSON#

A Series or DataFrame can be converted to a valid JSON string. Use to_json with optional parameters:

  • path_or_buf : the pathname or buffer to write the output This can be None in which case a JSON string is returned

  • orient :

    Series:
    • default is index

    • allowed values are {split, records, index}

    DataFrame:
    • default is columns

    • allowed values are {split, records, index, columns, values, table}

    The format of the JSON string

    split

    dict like {index -> [index], columns -> [columns], data -> [values]}

    records

    list like [{column -> value}, … , {column -> value}]

    index

    dict like {index -> {column -> value}}

    columns

    dict like {column -> {index -> value}}

    values

    just the values array

    table

    adhering to the JSON Table Schema

  • date_format : string, type of date conversion, ‘epoch’ for timestamp, ‘iso’ for ISO8601.

  • double_precision : The number of decimal places to use when encoding floating point values, default 10.

  • force_ascii : force encoded string to be ASCII, default True.

  • date_unit : The time unit to encode to, governs timestamp and ISO8601 precision. One of ‘s’, ‘ms’, ‘us’ or ‘ns’ for seconds, milliseconds, microseconds and nanoseconds respectively. Default ‘ms’.

  • default_handler : The handler to call if an object cannot otherwise be converted to a suitable format for JSON. Takes a single argument, which is the object to convert, and returns a serializable object.

  • lines : If records orient, then will write each record per line as json.

  • mode : string, writer mode when writing to path. ‘w’ for write, ‘a’ for append. Default ‘w’

Note NaN’s, NaT’s and None will be converted to null and datetime objects will be converted based on the date_format and date_unit parameters.

In [229]: dfj = pd.DataFrame(np.random.randn(5, 2), columns=list("AB"))

In [230]: json = dfj.to_json()

In [231]: json
Out[231]: '{"A":{"0":-0.1213062281,"1":0.6957746499,"2":0.9597255933,"3":-0.6199759194,"4":-0.7323393705},"B":{"0":-0.0978826728,"1":0.3417343559,"2":-1.1103361029,"3":0.1497483186,"4":0.6877383895}}'

Orient options#

There are a number of different options for the format of the resulting JSON file / string. Consider the following DataFrame and Series:

In [232]: dfjo = pd.DataFrame(
   .....:     dict(A=range(1, 4), B=range(4, 7), C=range(7, 10)),
   .....:     columns=list("ABC"),
   .....:     index=list("xyz"),
   .....: )
   .....: 

In [233]: dfjo
Out[233]: 
   A  B  C
x  1  4  7
y  2  5  8
z  3  6  9

In [234]: sjo = pd.Series(dict(x=15, y=16, z=17), name="D")

In [235]: sjo
Out[235]: 
x    15
y    16
z    17
Name: D, dtype: int64

Column oriented (the default for DataFrame) serializes the data as nested JSON objects with column labels acting as the primary index:

In [236]: dfjo.to_json(orient="columns")
Out[236]: '{"A":{"x":1,"y":2,"z":3},"B":{"x":4,"y":5,"z":6},"C":{"x":7,"y":8,"z":9}}'

# Not available for Series

Index oriented (the default for Series) similar to column oriented but the index labels are now primary:

In [237]: dfjo.to_json(orient="index")
Out[237]: '{"x":{"A":1,"B":4,"C":7},"y":{"A":2,"B":5,"C":8},"z":{"A":3,"B":6,"C":9}}'

In [238]: sjo.to_json(orient="index")
Out[238]: '{"x":15,"y":16,"z":17}'

Record oriented serializes the data to a JSON array of column -> value records, index labels are not included. This is useful for passing DataFrame data to plotting libraries, for example the JavaScript library d3.js:

In [239]: dfjo.to_json(orient="records")
Out[239]: '[{"A":1,"B":4,"C":7},{"A":2,"B":5,"C":8},{"A":3,"B":6,"C":9}]'

In [240]: sjo.to_json(orient="records")
Out[240]: '[15,16,17]'

Value oriented is a bare-bones option which serializes to nested JSON arrays of values only, column and index labels are not included:

In [241]: dfjo.to_json(orient="values")
Out[241]: '[[1,4,7],[2,5,8],[3,6,9]]'

# Not available for Series

Split oriented serializes to a JSON object containing separate entries for values, index and columns. Name is also included for Series:

In [242]: dfjo.to_json(orient="split")
Out[242]: '{"columns":["A","B","C"],"index":["x","y","z"],"data":[[1,4,7],[2,5,8],[3,6,9]]}'

In [243]: sjo.to_json(orient="split")
Out[243]: '{"name":"D","index":["x","y","z"],"data":[15,16,17]}'

Table oriented serializes to the JSON Table Schema, allowing for the preservation of metadata including but not limited to dtypes and index names.

Note

Any orient option that encodes to a JSON object will not preserve the ordering of index and column labels during round-trip serialization. If you wish to preserve label ordering use the split option as it uses ordered containers.

Date handling#

Writing in ISO date format:

In [244]: dfd = pd.DataFrame(np.random.randn(5, 2), columns=list("AB"))

In [245]: dfd["date"] = pd.Timestamp("20130101")

In [246]: dfd = dfd.sort_index(axis=1, ascending=False)

In [247]: json = dfd.to_json(date_format="iso")

In [248]: json
Out[248]: '{"date":{"0":"2013-01-01T00:00:00.000","1":"2013-01-01T00:00:00.000","2":"2013-01-01T00:00:00.000","3":"2013-01-01T00:00:00.000","4":"2013-01-01T00:00:00.000"},"B":{"0":0.403309524,"1":0.3016244523,"2":-1.3698493577,"3":1.4626960492,"4":-0.8265909164},"A":{"0":0.1764443426,"1":-0.1549507744,"2":-2.1798606054,"3":-0.9542078401,"4":-1.7431609117}}'

Writing in ISO date format, with microseconds:

In [249]: json = dfd.to_json(date_format="iso", date_unit="us")

In [250]: json
Out[250]: '{"date":{"0":"2013-01-01T00:00:00.000000","1":"2013-01-01T00:00:00.000000","2":"2013-01-01T00:00:00.000000","3":"2013-01-01T00:00:00.000000","4":"2013-01-01T00:00:00.000000"},"B":{"0":0.403309524,"1":0.3016244523,"2":-1.3698493577,"3":1.4626960492,"4":-0.8265909164},"A":{"0":0.1764443426,"1":-0.1549507744,"2":-2.1798606054,"3":-0.9542078401,"4":-1.7431609117}}'

Epoch timestamps, in seconds:

In [251]: json = dfd.to_json(date_format="epoch", date_unit="s")

In [252]: json
Out[252]: '{"date":{"0":1,"1":1,"2":1,"3":1,"4":1},"B":{"0":0.403309524,"1":0.3016244523,"2":-1.3698493577,"3":1.4626960492,"4":-0.8265909164},"A":{"0":0.1764443426,"1":-0.1549507744,"2":-2.1798606054,"3":-0.9542078401,"4":-1.7431609117}}'

Writing to a file, with a date index and a date column:

In [253]: dfj2 = dfj.copy()

In [254]: dfj2["date"] = pd.Timestamp("20130101")

In [255]: dfj2["ints"] = list(range(5))

In [256]: dfj2["bools"] = True

In [257]: dfj2.index = pd.date_range("20130101", periods=5)

In [258]: dfj2.to_json("test.json")

In [259]: with open("test.json") as fh:
   .....:     print(fh.read())
   .....: 
{"A":{"1356998400000":-0.1213062281,"1357084800000":0.6957746499,"1357171200000":0.9597255933,"1357257600000":-0.6199759194,"1357344000000":-0.7323393705},"B":{"1356998400000":-0.0978826728,"1357084800000":0.3417343559,"1357171200000":-1.1103361029,"1357257600000":0.1497483186,"1357344000000":0.6877383895},"date":{"1356998400000":1356,"1357084800000":1356,"1357171200000":1356,"1357257600000":1356,"1357344000000":1356},"ints":{"1356998400000":0,"1357084800000":1,"1357171200000":2,"1357257600000":3,"1357344000000":4},"bools":{"1356998400000":true,"1357084800000":true,"1357171200000":true,"1357257600000":true,"1357344000000":true}}

Fallback behavior#

If the JSON serializer cannot handle the container contents directly it will fall back in the following manner:

  • if the dtype is unsupported (e.g. np.complex_) then the default_handler, if provided, will be called for each value, otherwise an exception is raised.

  • if an object is unsupported it will attempt the following:

    • check if the object has defined a toDict method and call it. A toDict method should return a dict which will then be JSON serialized.

    • invoke the default_handler if one was provided.

    • convert the object to a dict by traversing its contents. However this will often fail with an OverflowError or give unexpected results.

In general the best approach for unsupported objects or dtypes is to provide a default_handler. For example:

>>> DataFrame([1.0, 2.0, complex(1.0, 2.0)]).to_json()  # raises
RuntimeError: Unhandled numpy dtype 15

can be dealt with by specifying a simple default_handler:

In [260]: pd.DataFrame([1.0, 2.0, complex(1.0, 2.0)]).to_json(default_handler=str)
Out[260]: '{"0":{"0":"(1+0j)","1":"(2+0j)","2":"(1+2j)"}}'

Reading JSON#

Reading a JSON string to pandas object can take a number of parameters. The parser will try to parse a DataFrame if typ is not supplied or is None. To explicitly force Series parsing, pass typ=series

  • filepath_or_buffer : a VALID JSON string or file handle / StringIO. The string could be a URL. Valid URL schemes include http, ftp, S3, and file. For file URLs, a host is expected. For instance, a local file could be file ://localhost/path/to/table.json

  • typ : type of object to recover (series or frame), default ‘frame’

  • orient :

    Series :
    • default is index

    • allowed values are {split, records, index}

    DataFrame
    • default is columns

    • allowed values are {split, records, index, columns, values, table}

    The format of the JSON string

    split

    dict like {index -> [index], columns -> [columns], data -> [values]}

    records

    list like [{column -> value}, … , {column -> value}]

    index

    dict like {index -> {column -> value}}

    columns

    dict like {column -> {index -> value}}

    values

    just the values array

    table

    adhering to the JSON Table Schema

  • dtype : if True, infer dtypes, if a dict of column to dtype, then use those, if False, then don’t infer dtypes at all, default is True, apply only to the data.

  • convert_axes : boolean, try to convert the axes to the proper dtypes, default is True

  • convert_dates : a list of columns to parse for dates; If True, then try to parse date-like columns, default is True.

  • keep_default_dates : boolean, default True. If parsing dates, then parse the default date-like columns.

  • precise_float : boolean, default False. Set to enable usage of higher precision (strtod) function when decoding string to double values. Default (False) is to use fast but less precise builtin functionality.

  • date_unit : string, the timestamp unit to detect if converting dates. Default None. By default the timestamp precision will be detected, if this is not desired then pass one of ‘s’, ‘ms’, ‘us’ or ‘ns’ to force timestamp precision to seconds, milliseconds, microseconds or nanoseconds respectively.

  • lines : reads file as one json object per line.

  • encoding : The encoding to use to decode py3 bytes.

  • chunksize : when used in combination with lines=True, return a pandas.api.typing.JsonReader which reads in chunksize lines per iteration.

  • engine: Either "ujson", the built-in JSON parser, or "pyarrow" which dispatches to pyarrow’s pyarrow.json.read_json. The "pyarrow" is only available when lines=True

The parser will raise one of ValueError/TypeError/AssertionError if the JSON is not parseable.

If a non-default orient was used when encoding to JSON be sure to pass the same option here so that decoding produces sensible results, see Orient Options for an overview.

Data conversion#

The default of convert_axes=True, dtype=True, and convert_dates=True will try to parse the axes, and all of the data into appropriate types, including dates. If you need to override specific dtypes, pass a dict to dtype. convert_axes should only be set to False if you need to preserve string-like numbers (e.g. ‘1’, ‘2’) in an axes.

Note

Large integer values may be converted to dates if convert_dates=True and the data and / or column labels appear ‘date-like’. The exact threshold depends on the date_unit specified. ‘date-like’ means that the column label meets one of the following criteria:

  • it ends with '_at'

  • it ends with '_time'

  • it begins with 'timestamp'

  • it is 'modified'

  • it is 'date'

Warning

When reading JSON data, automatic coercing into dtypes has some quirks:

  • an index can be reconstructed in a different order from serialization, that is, the returned order is not guaranteed to be the same as before serialization

  • a column that was float data will be converted to integer if it can be done safely, e.g. a column of 1.

  • bool columns will be converted to integer on reconstruction

Thus there are times where you may want to specify specific dtypes via the dtype keyword argument.

Reading from a JSON string:

In [261]: from io import StringIO

In [262]: pd.read_json(StringIO(json))
Out[262]: 
   date         B         A
0     1  0.403310  0.176444
1     1  0.301624 -0.154951
2     1 -1.369849 -2.179861
3     1  1.462696 -0.954208
4     1 -0.826591 -1.743161

Reading from a file:

In [263]: pd.read_json("test.json")
Out[263]: 
                   A         B  date  ints  bools
2013-01-01 -0.121306 -0.097883  1356     0   True
2013-01-02  0.695775  0.341734  1356     1   True
2013-01-03  0.959726 -1.110336  1356     2   True
2013-01-04 -0.619976  0.149748  1356     3   True
2013-01-05 -0.732339  0.687738  1356     4   True

Don’t convert any data (but still convert axes and dates):

In [264]: pd.read_json("test.json", dtype=object).dtypes
Out[264]: 
A        object
B        object
date     object
ints     object
bools    object
dtype: object

Specify dtypes for conversion:

In [265]: pd.read_json("test.json", dtype={"A": "float32", "bools": "int8"}).dtypes
Out[265]: 
A        float32
B        float64
date       int64
ints       int64
bools       int8
dtype: object

Preserve string indices:

In [266]: from io import StringIO

In [267]: si = pd.DataFrame(
   .....:     np.zeros((4, 4)), columns=list(range(4)), index=[str(i) for i in range(4)]
   .....: )
   .....: 

In [268]: si
Out[268]: 
     0    1    2    3
0  0.0  0.0  0.0  0.0
1  0.0  0.0  0.0  0.0
2  0.0  0.0  0.0  0.0
3  0.0  0.0  0.0  0.0

In [269]: si.index
Out[269]: Index(['0', '1', '2', '3'], dtype='object')

In [270]: si.columns
Out[270]: Index([0, 1, 2, 3], dtype='int64')

In [271]: json = si.to_json()

In [272]: sij = pd.read_json(StringIO(json), convert_axes=False)

In [273]: sij
Out[273]: 
   0  1  2  3
0  0  0  0  0
1  0  0  0  0
2  0  0  0  0
3  0  0  0  0

In [274]: sij.index
Out[274]: Index(['0', '1', '2', '3'], dtype='object')

In [275]: sij.columns
Out[275]: Index(['0', '1', '2', '3'], dtype='object')

Dates written in nanoseconds need to be read back in nanoseconds:

In [276]: from io import StringIO

In [277]: json = dfj2.to_json(date_unit="ns")

# Try to parse timestamps as milliseconds -> Won't Work
In [278]: dfju = pd.read_json(StringIO(json), date_unit="ms")

In [279]: dfju
Out[279]: 
                            A         B        date  ints  bools
1356998400000000000 -0.121306 -0.097883  1356998400     0   True
1357084800000000000  0.695775  0.341734  1356998400     1   True
1357171200000000000  0.959726 -1.110336  1356998400     2   True
1357257600000000000 -0.619976  0.149748  1356998400     3   True
1357344000000000000 -0.732339  0.687738  1356998400     4   True

# Let pandas detect the correct precision
In [280]: dfju = pd.read_json(StringIO(json))

In [281]: dfju
Out[281]: 
                   A         B       date  ints  bools
2013-01-01 -0.121306 -0.097883 2013-01-01     0   True
2013-01-02  0.695775  0.341734 2013-01-01     1   True
2013-01-03  0.959726 -1.110336 2013-01-01     2   True
2013-01-04 -0.619976  0.149748 2013-01-01     3   True
2013-01-05 -0.732339  0.687738 2013-01-01     4   True

# Or specify that all timestamps are in nanoseconds
In [282]: dfju = pd.read_json(StringIO(json), date_unit="ns")

In [283]: dfju
Out[283]: 
                   A         B        date  ints  bools
2013-01-01 -0.121306 -0.097883  1356998400     0   True
2013-01-02  0.695775  0.341734  1356998400     1   True
2013-01-03  0.959726 -1.110336  1356998400     2   True
2013-01-04 -0.619976  0.149748  1356998400     3   True
2013-01-05 -0.732339  0.687738  1356998400     4   True

By setting the dtype_backend argument you can control the default dtypes used for the resulting DataFrame.

In [284]: data = (
   .....:  '{"a":{"0":1,"1":3},"b":{"0":2.5,"1":4.5},"c":{"0":true,"1":false},"d":{"0":"a","1":"b"},'
   .....:  '"e":{"0":null,"1":6.0},"f":{"0":null,"1":7.5},"g":{"0":null,"1":true},"h":{"0":null,"1":"a"},'
   .....:  '"i":{"0":"12-31-2019","1":"12-31-2019"},"j":{"0":null,"1":null}}'
   .....: )
   .....: 

In [285]: df = pd.read_json(StringIO(data), dtype_backend="pyarrow")

In [286]: df
Out[286]: 
   a    b      c  d     e     f     g     h           i     j
0  1  2.5   True  a  <NA>  <NA>  <NA>  <NA>  12-31-2019  None
1  3  4.5  False  b     6   7.5  True     a  12-31-2019  None

In [287]: df.dtypes
Out[287]: 
a     int64[pyarrow]
b    double[pyarrow]
c      bool[pyarrow]
d    string[pyarrow]
e     int64[pyarrow]
f    double[pyarrow]
g      bool[pyarrow]
h    string[pyarrow]
i    string[pyarrow]
j             object
dtype: object

Normalization#

pandas provides a utility function to take a dict or list of dicts and normalize this semi-structured data into a flat table.

In [288]: data = [
   .....:     {"id": 1, "name": {"first": "Coleen", "last": "Volk"}},
   .....:     {"name": {"given": "Mark", "family": "Regner"}},
   .....:     {"id": 2, "name": "Faye Raker"},
   .....: ]
   .....: 

In [289]: pd.json_normalize(data)
Out[289]: 
    id name.first name.last name.given name.family        name
0  1.0     Coleen      Volk        NaN         NaN         NaN
1  NaN        NaN       NaN       Mark      Regner         NaN
2  2.0        NaN       NaN        NaN         NaN  Faye Raker
In [290]: data = [
   .....:     {
   .....:         "state": "Florida",
   .....:         "shortname": "FL",
   .....:         "info": {"governor": "Rick Scott"},
   .....:         "county": [
   .....:             {"name": "Dade", "population": 12345},
   .....:             {"name": "Broward", "population": 40000},
   .....:             {"name": "Palm Beach", "population": 60000},
   .....:         ],
   .....:     },
   .....:     {
   .....:         "state": "Ohio",
   .....:         "shortname": "OH",
   .....:         "info": {"governor": "John Kasich"},
   .....:         "county": [
   .....:             {"name": "Summit", "population": 1234},
   .....:             {"name": "Cuyahoga", "population": 1337},
   .....:         ],
   .....:     },
   .....: ]
   .....: 

In [291]: pd.json_normalize(data, "county", ["state", "shortname", ["info", "governor"]])
Out[291]: 
         name  population    state shortname info.governor
0        Dade       12345  Florida        FL    Rick Scott
1     Broward       40000  Florida        FL    Rick Scott
2  Palm Beach       60000  Florida        FL    Rick Scott
3      Summit        1234     Ohio        OH   John Kasich
4    Cuyahoga        1337     Ohio        OH   John Kasich

The max_level parameter provides more control over which level to end normalization. With max_level=1 the following snippet normalizes until 1st nesting level of the provided dict.

In [292]: data = [
   .....:     {
   .....:         "CreatedBy": {"Name": "User001"},
   .....:         "Lookup": {
   .....:             "TextField": "Some text",
   .....:             "UserField": {"Id": "ID001", "Name": "Name001"},
   .....:         },
   .....:         "Image": {"a": "b"},
   .....:     }
   .....: ]
   .....: 

In [293]: pd.json_normalize(data, max_level=1)
Out[293]: 
  CreatedBy.Name Lookup.TextField                    Lookup.UserField Image.a
0        User001        Some text  {'Id': 'ID001', 'Name': 'Name001'}       b

Line delimited json#

pandas is able to read and write line-delimited json files that are common in data processing pipelines using Hadoop or Spark.

For line-delimited json files, pandas can also return an iterator which reads in chunksize lines at a time. This can be useful for large files or to read from a stream.

In [294]: from io import StringIO

In [295]: jsonl = """
   .....:     {"a": 1, "b": 2}
   .....:     {"a": 3, "b": 4}
   .....: """
   .....: 

In [296]: df = pd.read_json(StringIO(jsonl), lines=True)

In [297]: df
Out[297]: 
   a  b
0  1  2
1  3  4

In [298]: df.to_json(orient="records", lines=True)
Out[298]: '{"a":1,"b":2}\n{"a":3,"b":4}\n'

# reader is an iterator that returns ``chunksize`` lines each iteration
In [299]: with pd.read_json(StringIO(jsonl), lines=True, chunksize=1) as reader:
   .....:     reader
   .....:     for chunk in reader:
   .....:         print(chunk)
   .....: 
Empty DataFrame
Columns: []
Index: []
   a  b
0  1  2
   a  b
1  3  4

Line-limited json can also be read using the pyarrow reader by specifying engine="pyarrow".

In [300]: from io import BytesIO

In [301]: df = pd.read_json(BytesIO(jsonl.encode()), lines=True, engine="pyarrow")

In [302]: df
Out[302]: 
   a  b
0  1  2
1  3  4

New in version 2.0.0.

Table schema#

Table Schema is a spec for describing tabular datasets as a JSON object. The JSON includes information on the field names, types, and other attributes. You can use the orient table to build a JSON string with two fields, schema and data.

In [303]: df = pd.DataFrame(
   .....:     {
   .....:         "A": [1, 2, 3],
   .....:         "B": ["a", "b", "c"],
   .....:         "C": pd.date_range("2016-01-01", freq="d", periods=3),
   .....:     },
   .....:     index=pd.Index(range(3), name="idx"),
   .....: )
   .....: 

In [304]: df
Out[304]: 
     A  B          C
idx                 
0    1  a 2016-01-01
1    2  b 2016-01-02
2    3  c 2016-01-03

In [305]: df.to_json(orient="table", date_format="iso")
Out[305]: '{"schema":{"fields":[{"name":"idx","type":"integer"},{"name":"A","type":"integer"},{"name":"B","type":"string"},{"name":"C","type":"datetime"}],"primaryKey":["idx"],"pandas_version":"1.4.0"},"data":[{"idx":0,"A":1,"B":"a","C":"2016-01-01T00:00:00.000"},{"idx":1,"A":2,"B":"b","C":"2016-01-02T00:00:00.000"},{"idx":2,"A":3,"B":"c","C":"2016-01-03T00:00:00.000"}]}'

The schema field contains the fields key, which itself contains a list of column name to type pairs, including the Index or MultiIndex (see below for a list of types). The schema field also contains a primaryKey field if the (Multi)index is unique.

The second field, data, contains the serialized data with the records orient. The index is included, and any datetimes are ISO 8601 formatted, as required by the Table Schema spec.

The full list of types supported are described in the Table Schema spec. This table shows the mapping from pandas types:

pandas type

Table Schema type

int64

integer

float64

number

bool

boolean

datetime64[ns]

datetime

timedelta64[ns]

duration

categorical

any

object

str

A few notes on the generated table schema:

  • The schema object contains a pandas_version field. This contains the version of pandas’ dialect of the schema, and will be incremented with each revision.

  • All dates are converted to UTC when serializing. Even timezone naive values, which are treated as UTC with an offset of 0.

    In [306]: from pandas.io.json import build_table_schema
    
    In [307]: s = pd.Series(pd.date_range("2016", periods=4))
    
    In [308]: build_table_schema(s)
    Out[308]: 
    {'fields': [{'name': 'index', 'type': 'integer'},
      {'name': 'values', 'type': 'datetime'}],
     'primaryKey': ['index'],
     'pandas_version': '1.4.0'}
    
  • datetimes with a timezone (before serializing), include an additional field tz with the time zone name (e.g. 'US/Central').

    In [309]: s_tz = pd.Series(pd.date_range("2016", periods=12, tz="US/Central"))
    
    In [310]: build_table_schema(s_tz)
    Out[310]: 
    {'fields': [{'name': 'index', 'type': 'integer'},
      {'name': 'values', 'type': 'datetime', 'tz': 'US/Central'}],
     'primaryKey': ['index'],
     'pandas_version': '1.4.0'}
    
  • Periods are converted to timestamps before serialization, and so have the same behavior of being converted to UTC. In addition, periods will contain and additional field freq with the period’s frequency, e.g. 'A-DEC'.

    In [311]: s_per = pd.Series(1, index=pd.period_range("2016", freq="A-DEC", periods=4))
    
    In [312]: build_table_schema(s_per)
    Out[312]: 
    {'fields': [{'name': 'index', 'type': 'datetime', 'freq': 'A-DEC'},
      {'name': 'values', 'type': 'integer'}],
     'primaryKey': ['index'],
     'pandas_version': '1.4.0'}
    
  • Categoricals use the any type and an enum constraint listing the set of possible values. Additionally, an ordered field is included:

    In [313]: s_cat = pd.Series(pd.Categorical(["a", "b", "a"]))
    
    In [314]: build_table_schema(s_cat)
    Out[314]: 
    {'fields': [{'name': 'index', 'type': 'integer'},
      {'name': 'values',
       'type': 'any',
       'constraints': {'enum': ['a', 'b']},
       'ordered': False}],
     'primaryKey': ['index'],
     'pandas_version': '1.4.0'}
    
  • A primaryKey field, containing an array of labels, is included if the index is unique:

    In [315]: s_dupe = pd.Series([1, 2], index=[1, 1])
    
    In [316]: build_table_schema(s_dupe)
    Out[316]: 
    {'fields': [{'name': 'index', 'type': 'integer'},
      {'name': 'values', 'type': 'integer'}],
     'pandas_version': '1.4.0'}
    
  • The primaryKey behavior is the same with MultiIndexes, but in this case the primaryKey is an array:

    In [317]: s_multi = pd.Series(1, index=pd.MultiIndex.from_product([("a", "b"), (0, 1)]))
    
    In [318]: build_table_schema(s_multi)
    Out[318]: 
    {'fields': [{'name': 'level_0', 'type': 'string'},
      {'name': 'level_1', 'type': 'integer'},
      {'name': 'values', 'type': 'integer'}],
     'primaryKey': FrozenList(['level_0', 'level_1']),
     'pandas_version': '1.4.0'}
    
  • The default naming roughly follows these rules:

    • For series, the object.name is used. If that’s none, then the name is values

    • For DataFrames, the stringified version of the column name is used

    • For Index (not MultiIndex), index.name is used, with a fallback to index if that is None.

    • For MultiIndex, mi.names is used. If any level has no name, then level_<i> is used.

read_json also accepts orient='table' as an argument. This allows for the preservation of metadata such as dtypes and index names in a round-trippable manner.

In [319]: df = pd.DataFrame(
   .....:     {
   .....:         "foo": [1, 2, 3, 4],
   .....:         "bar": ["a", "b", "c", "d"],
   .....:         "baz": pd.date_range("2018-01-01", freq="d", periods=4),
   .....:         "qux": pd.Categorical(["a", "b", "c", "c"]),
   .....:     },
   .....:     index=pd.Index(range(4), name="idx"),
   .....: )
   .....: 

In [320]: df
Out[320]: 
     foo bar        baz qux
idx                        
0      1   a 2018-01-01   a
1      2   b 2018-01-02   b
2      3   c 2018-01-03   c
3      4   d 2018-01-04   c

In [321]: df.dtypes
Out[321]: 
foo             int64
bar            object
baz    datetime64[ns]
qux          category
dtype: object

In [322]: df.to_json("test.json", orient="table")

In [323]: new_df = pd.read_json("test.json", orient="table")

In [324]: new_df
Out[324]: 
     foo bar        baz qux
idx                        
0      1   a 2018-01-01   a
1      2   b 2018-01-02   b
2      3   c 2018-01-03   c
3      4   d 2018-01-04   c

In [325]: new_df.dtypes
Out[325]: 
foo             int64
bar            object
baz    datetime64[ns]
qux          category
dtype: object

Please note that the literal string ‘index’ as the name of an Index is not round-trippable, nor are any names beginning with 'level_' within a MultiIndex. These are used by default in DataFrame.to_json() to indicate missing values and the subsequent read cannot distinguish the intent.

In [326]: df.index.name = "index"

In [327]: df.to_json("test.json", orient="table")

In [328]: new_df = pd.read_json("test.json", orient="table")

In [329]: print(new_df.index.name)
None

When using orient='table' along with user-defined ExtensionArray, the generated schema will contain an additional extDtype key in the respective fields element. This extra key is not standard but does enable JSON roundtrips for extension types (e.g. read_json(df.to_json(orient="table"), orient="table")).

The extDtype key carries the name of the extension, if you have properly registered the ExtensionDtype, pandas will use said name to perform a lookup into the registry and re-convert the serialized data into your custom dtype.

HTML#

Reading HTML content#

Warning

We highly encourage you to read the HTML Table Parsing gotchas below regarding the issues surrounding the BeautifulSoup4/html5lib/lxml parsers.

The top-level read_html() function can accept an HTML string/file/URL and will parse HTML tables into list of pandas DataFrames. Let’s look at a few examples.

Note

read_html returns a list of DataFrame objects, even if there is only a single table contained in the HTML content.

Read a URL with no options:

In [320]: "https://www.fdic.gov/resources/resolutions/bank-failures/failed-bank-list"
In [321]: pd.read_html(url)
Out[321]:
[                         Bank NameBank           CityCity StateSt  ...              Acquiring InstitutionAI Closing DateClosing FundFund
 0                    Almena State Bank             Almena      KS  ...                          Equity Bank    October 23, 2020    10538
 1           First City Bank of Florida  Fort Walton Beach      FL  ...            United Fidelity Bank, fsb    October 16, 2020    10537
 2                 The First State Bank      Barboursville      WV  ...                       MVB Bank, Inc.       April 3, 2020    10536
 3                   Ericson State Bank            Ericson      NE  ...           Farmers and Merchants Bank   February 14, 2020    10535
 4     City National Bank of New Jersey             Newark      NJ  ...                      Industrial Bank    November 1, 2019    10534
 ..                                 ...                ...     ...  ...                                  ...                 ...      ...
 558                 Superior Bank, FSB           Hinsdale      IL  ...                Superior Federal, FSB       July 27, 2001     6004
 559                Malta National Bank              Malta      OH  ...                    North Valley Bank         May 3, 2001     4648
 560    First Alliance Bank & Trust Co.         Manchester      NH  ...  Southern New Hampshire Bank & Trust    February 2, 2001     4647
 561  National State Bank of Metropolis         Metropolis      IL  ...              Banterra Bank of Marion   December 14, 2000     4646
 562                   Bank of Honolulu           Honolulu      HI  ...                   Bank of the Orient    October 13, 2000     4645

 [563 rows x 7 columns]]

Note

The data from the above URL changes every Monday so the resulting data above may be slightly different.

Read a URL while passing headers alongside the HTTP request:

In [322]: url = 'https://www.sump.org/notes/request/' # HTTP request reflector
In [323]: pd.read_html(url)
Out[323]:
[                   0                    1
 0     Remote Socket:  51.15.105.256:51760
 1  Protocol Version:             HTTP/1.1
 2    Request Method:                  GET
 3       Request URI:      /notes/request/
 4     Request Query:                  NaN,
 0   Accept-Encoding:             identity
 1              Host:         www.sump.org
 2        User-Agent:    Python-urllib/3.8
 3        Connection:                close]
In [324]: headers = {
In [325]:    'User-Agent':'Mozilla Firefox v14.0',
In [326]:    'Accept':'application/json',
In [327]:    'Connection':'keep-alive',
In [328]:    'Auth':'Bearer 2*/f3+fe68df*4'
In [329]: }
In [340]: pd.read_html(url, storage_options=headers)
Out[340]:
[                   0                    1
 0     Remote Socket:  51.15.105.256:51760
 1  Protocol Version:             HTTP/1.1
 2    Request Method:                  GET
 3       Request URI:      /notes/request/
 4     Request Query:                  NaN,
 0        User-Agent: Mozilla Firefox v14.0
 1    AcceptEncoding:   gzip,  deflate,  br
 2            Accept:      application/json
 3        Connection:             keep-alive
 4              Auth:  Bearer 2*/f3+fe68df*4]

Note

We see above that the headers we passed are reflected in the HTTP request.

Read in the content of the file from the above URL and pass it to read_html as a string:

In [330]: html_str = """
   .....:          <table>
   .....:              <tr>
   .....:                  <th>A</th>
   .....:                  <th colspan="1">B</th>
   .....:                  <th rowspan="1">C</th>
   .....:              </tr>
   .....:              <tr>
   .....:                  <td>a</td>
   .....:                  <td>b</td>
   .....:                  <td>c</td>
   .....:              </tr>
   .....:          </table>
   .....:      """
   .....: 

In [331]: with open("tmp.html", "w") as f:
   .....:     f.write(html_str)
   .....: 

In [332]: df = pd.read_html("tmp.html")

In [333]: df[0]
Out[333]: 
   A  B  C
0  a  b  c

You can even pass in an instance of StringIO if you so desire:

In [334]: dfs = pd.read_html(StringIO(html_str))

In [335]: dfs[0]
Out[335]: 
   A  B  C
0  a  b  c

Note

The following examples are not run by the IPython evaluator due to the fact that having so many network-accessing functions slows down the documentation build. If you spot an error or an example that doesn’t run, please do not hesitate to report it over on pandas GitHub issues page.

Read a URL and match a table that contains specific text:

match = "Metcalf Bank"
df_list = pd.read_html(url, match=match)

Specify a header row (by default <th> or <td> elements located within a <thead> are used to form the column index, if multiple rows are contained within <thead> then a MultiIndex is created); if specified, the header row is taken from the data minus the parsed header elements (<th> elements).

dfs = pd.read_html(url, header=0)

Specify an index column:

dfs = pd.read_html(url, index_col=0)

Specify a number of rows to skip:

dfs = pd.read_html(url, skiprows=0)

Specify a number of rows to skip using a list (range works as well):

dfs = pd.read_html(url, skiprows=range(2))

Specify an HTML attribute:

dfs1 = pd.read_html(url, attrs={"id": "table"})
dfs2 = pd.read_html(url, attrs={"class": "sortable"})
print(np.array_equal(dfs1[0], dfs2[0]))  # Should be True

Specify values that should be converted to NaN:

dfs = pd.read_html(url, na_values=["No Acquirer"])

Specify whether to keep the default set of NaN values:

dfs = pd.read_html(url, keep_default_na=False)

Specify converters for columns. This is useful for numerical text data that has leading zeros. By default columns that are numerical are cast to numeric types and the leading zeros are lost. To avoid this, we can convert these columns to strings.

url_mcc = "https://en.wikipedia.org/wiki/Mobile_country_code"
dfs = pd.read_html(
    url_mcc,
    match="Telekom Albania",
    header=0,
    converters={"MNC": str},
)

Use some combination of the above:

dfs = pd.read_html(url, match="Metcalf Bank", index_col=0)

Read in pandas to_html output (with some loss of floating point precision):

df = pd.DataFrame(np.random.randn(2, 2))
s = df.to_html(float_format="{0:.40g}".format)
dfin = pd.read_html(s, index_col=0)

The lxml backend will raise an error on a failed parse if that is the only parser you provide. If you only have a single parser you can provide just a string, but it is considered good practice to pass a list with one string if, for example, the function expects a sequence of strings. You may use:

dfs = pd.read_html(url, "Metcalf Bank", index_col=0, flavor=["lxml"])

Or you could pass flavor='lxml' without a list:

dfs = pd.read_html(url, "Metcalf Bank", index_col=0, flavor="lxml")

However, if you have bs4 and html5lib installed and pass None or ['lxml', 'bs4'] then the parse will most likely succeed. Note that as soon as a parse succeeds, the function will return.

dfs = pd.read_html(url, "Metcalf Bank", index_col=0, flavor=["lxml", "bs4"])

Links can be extracted from cells along with the text using extract_links="all".

In [336]: html_table = """
   .....: <table>
   .....:   <tr>
   .....:     <th>GitHub</th>
   .....:   </tr>
   .....:   <tr>
   .....:     <td><a href="https://github.com/pandas-dev/pandas">pandas</a></td>
   .....:   </tr>
   .....: </table>
   .....: """
   .....: 

In [337]: df = pd.read_html(
   .....:     StringIO(html_table),
   .....:     extract_links="all"
   .....: )[0]
   .....: 

In [338]: df
Out[338]: 
                                   (GitHub, None)
0  (pandas, https://github.com/pandas-dev/pandas)

In [339]: df[("GitHub", None)]
Out[339]: 
0    (pandas, https://github.com/pandas-dev/pandas)
Name: (GitHub, None), dtype: object

In [340]: df[("GitHub", None)].str[1]
Out[340]: 
0    https://github.com/pandas-dev/pandas
Name: (GitHub, None), dtype: object

New in version 1.5.0.

Writing to HTML files#

DataFrame objects have an instance method to_html which renders the contents of the DataFrame as an HTML table. The function arguments are as in the method to_string described above.

Note

Not all of the possible options for DataFrame.to_html are shown here for brevity’s sake. See to_html() for the full set of options.

Note

In an HTML-rendering supported environment like a Jupyter Notebook, display(HTML(...))` will render the raw HTML into the environment.

In [341]: from IPython.display import display, HTML

In [342]: df = pd.DataFrame(np.random.randn(2, 2))

In [343]: df
Out[343]: 
          0         1
0 -0.345352  1.314232
1  0.690579  0.995761

In [344]: html = df.to_html()

In [345]: print(html)  # raw html
<table border="1" class="dataframe">
  <thead>
    <tr style="text-align: right;">
      <th></th>
      <th>0</th>
      <th>1</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <th>0</th>
      <td>-0.345352</td>
      <td>1.314232</td>
    </tr>
    <tr>
      <th>1</th>
      <td>0.690579</td>
      <td>0.995761</td>
    </tr>
  </tbody>
</table>

In [346]: display(HTML(html))
<IPython.core.display.HTML object>

The columns argument will limit the columns shown:

In [347]: html = df.to_html(columns=[0])

In [348]: print(html)
<table border="1" class="dataframe">
  <thead>
    <tr style="text-align: right;">
      <th></th>
      <th>0</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <th>0</th>
      <td>-0.345352</td>
    </tr>
    <tr>
      <th>1</th>
      <td>0.690579</td>
    </tr>
  </tbody>
</table>

In [349]: display(HTML(html))
<IPython.core.display.HTML object>

float_format takes a Python callable to control the precision of floating point values:

In [350]: html = df.to_html(float_format="{0:.10f}".format)

In [351]: print(html)
<table border="1" class="dataframe">
  <thead>
    <tr style="text-align: right;">
      <th></th>
      <th>0</th>
      <th>1</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <th>0</th>
      <td>-0.3453521949</td>
      <td>1.3142323796</td>
    </tr>
    <tr>
      <th>1</th>
      <td>0.6905793352</td>
      <td>0.9957609037</td>
    </tr>
  </tbody>
</table>

In [352]: display(HTML(html))
<IPython.core.display.HTML object>

bold_rows will make the row labels bold by default, but you can turn that off:

In [353]: html = df.to_html(bold_rows=False)

In [354]: print(html)
<table border="1" class="dataframe">
  <thead>
    <tr style="text-align: right;">
      <th></th>
      <th>0</th>
      <th>1</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>0</td>
      <td>-0.345352</td>
      <td>1.314232</td>
    </tr>
    <tr>
      <td>1</td>
      <td>0.690579</td>
      <td>0.995761</td>
    </tr>
  </tbody>
</table>

In [355]: display(HTML(html))
<IPython.core.display.HTML object>

The classes argument provides the ability to give the resulting HTML table CSS classes. Note that these classes are appended to the existing 'dataframe' class.

In [356]: print(df.to_html(classes=["awesome_table_class", "even_more_awesome_class"]))
<table border="1" class="dataframe awesome_table_class even_more_awesome_class">
  <thead>
    <tr style="text-align: right;">
      <th></th>
      <th>0</th>
      <th>1</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <th>0</th>
      <td>-0.345352</td>
      <td>1.314232</td>
    </tr>
    <tr>
      <th>1</th>
      <td>0.690579</td>
      <td>0.995761</td>
    </tr>
  </tbody>
</table>

The render_links argument provides the ability to add hyperlinks to cells that contain URLs.

In [357]: url_df = pd.DataFrame(
   .....:     {
   .....:         "name": ["Python", "pandas"],
   .....:         "url": ["https://www.python.org/", "https://pandas.pydata.org"],
   .....:     }
   .....: )
   .....: 

In [358]: html = url_df.to_html(render_links=True)

In [359]: print(html)
<table border="1" class="dataframe">
  <thead>
    <tr style="text-align: right;">
      <th></th>
      <th>name</th>
      <th>url</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <th>0</th>
      <td>Python</td>
      <td><a href="https://www.python.org/" target="_blank">https://www.python.org/</a></td>
    </tr>
    <tr>
      <th>1</th>
      <td>pandas</td>
      <td><a href="https://pandas.pydata.org" target="_blank">https://pandas.pydata.org</a></td>
    </tr>
  </tbody>
</table>

In [360]: display(HTML(html))
<IPython.core.display.HTML object>

Finally, the escape argument allows you to control whether the “<”, “>” and “&” characters escaped in the resulting HTML (by default it is True). So to get the HTML without escaped characters pass escape=False

In [361]: df = pd.DataFrame({"a": list("&<>"), "b": np.random.randn(3)})

Escaped:

In [362]: html = df.to_html()

In [363]: print(html)
<table border="1" class="dataframe">
  <thead>
    <tr style="text-align: right;">
      <th></th>
      <th>a</th>
      <th>b</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <th>0</th>
      <td>&amp;</td>
      <td>2.396780</td>
    </tr>
    <tr>
      <th>1</th>
      <td>&lt;</td>
      <td>0.014871</td>
    </tr>
    <tr>
      <th>2</th>
      <td>&gt;</td>
      <td>3.357427</td>
    </tr>
  </tbody>
</table>

In [364]: display(HTML(html))
<IPython.core.display.HTML object>

Not escaped:

In [365]: html = df.to_html(escape=False)

In [366]: print(html)
<table border="1" class="dataframe">
  <thead>
    <tr style="text-align: right;">
      <th></th>
      <th>a</th>
      <th>b</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <th>0</th>
      <td>&</td>
      <td>2.396780</td>
    </tr>
    <tr>
      <th>1</th>
      <td><</td>
      <td>0.014871</td>
    </tr>
    <tr>
      <th>2</th>
      <td>></td>
      <td>3.357427</td>
    </tr>
  </tbody>
</table>

In [367]: display(HTML(html))
<IPython.core.display.HTML object>

Note

Some browsers may not show a difference in the rendering of the previous two HTML tables.

HTML Table Parsing Gotchas#

There are some versioning issues surrounding the libraries that are used to parse HTML tables in the top-level pandas io function read_html.

Issues with lxml

  • Benefits

    • lxml is very fast.

    • lxml requires Cython to install correctly.

  • Drawbacks

    • lxml does not make any guarantees about the results of its parse unless it is given strictly valid markup.

    • In light of the above, we have chosen to allow you, the user, to use the lxml backend, but this backend will use html5lib if lxml fails to parse

    • It is therefore highly recommended that you install both BeautifulSoup4 and html5lib, so that you will still get a valid result (provided everything else is valid) even if lxml fails.

Issues with BeautifulSoup4 using lxml as a backend

  • The above issues hold here as well since BeautifulSoup4 is essentially just a wrapper around a parser backend.

Issues with BeautifulSoup4 using html5lib as a backend

  • Benefits

    • html5lib is far more lenient than lxml and consequently deals with real-life markup in a much saner way rather than just, e.g., dropping an element without notifying you.

    • html5lib generates valid HTML5 markup from invalid markup automatically. This is extremely important for parsing HTML tables, since it guarantees a valid document. However, that does NOT mean that it is “correct”, since the process of fixing markup does not have a single definition.

    • html5lib is pure Python and requires no additional build steps beyond its own installation.

  • Drawbacks

    • The biggest drawback to using html5lib is that it is slow as molasses. However consider the fact that many tables on the web are not big enough for the parsing algorithm runtime to matter. It is more likely that the bottleneck will be in the process of reading the raw text from the URL over the web, i.e., IO (input-output). For very large tables, this might not be true.

LaTeX#

New in version 1.3.0.

Currently there are no methods to read from LaTeX, only output methods.

Writing to LaTeX files#

Note

DataFrame and Styler objects currently have a to_latex method. We recommend using the Styler.to_latex() method over DataFrame.to_latex() due to the former’s greater flexibility with conditional styling, and the latter’s possible future deprecation.

Review the documentation for Styler.to_latex, which gives examples of conditional styling and explains the operation of its keyword arguments.

For simple application the following pattern is sufficient.

In [368]: df = pd.DataFrame([[1, 2], [3, 4]], index=["a", "b"], columns=["c", "d"])

In [369]: print(df.style.to_latex())
\begin{tabular}{lrr}
 & c & d \\
a & 1 & 2 \\
b & 3 & 4 \\
\end{tabular}

To format values before output, chain the Styler.format method.

In [370]: print(df.style.format("€ {}").to_latex())
\begin{tabular}{lrr}
 & c & d \\
a & € 1 & € 2 \\
b & € 3 & € 4 \\
\end{tabular}

XML#

Reading XML#

New in version 1.3.0.

The top-level read_xml() function can accept an XML string/file/URL and will parse nodes and attributes into a pandas DataFrame.

Note

Since there is no standard XML structure where design types can vary in many ways, read_xml works best with flatter, shallow versions. If an XML document is deeply nested, use the stylesheet feature to transform XML into a flatter version.

Let’s look at a few examples.

Read an XML string:

In [371]: from io import StringIO

In [372]: xml = """<?xml version="1.0" encoding="UTF-8"?>
   .....: <bookstore>
   .....:   <book category="cooking">
   .....:     <title lang="en">Everyday Italian</title>
   .....:     <author>Giada De Laurentiis</author>
   .....:     <year>2005</year>
   .....:     <price>30.00</price>
   .....:   </book>
   .....:   <book category="children">
   .....:     <title lang="en">Harry Potter</title>
   .....:     <author>J K. Rowling</author>
   .....:     <year>2005</year>
   .....:     <price>29.99</price>
   .....:   </book>
   .....:   <book category="web">
   .....:     <title lang="en">Learning XML</title>
   .....:     <author>Erik T. Ray</author>
   .....:     <year>2003</year>
   .....:     <price>39.95</price>
   .....:   </book>
   .....: </bookstore>"""
   .....: 

In [373]: df = pd.read_xml(StringIO(xml))

In [374]: df
Out[374]: 
   category             title               author  year  price
0   cooking  Everyday Italian  Giada De Laurentiis  2005  30.00
1  children      Harry Potter         J K. Rowling  2005  29.99
2       web      Learning XML          Erik T. Ray  2003  39.95

Read a URL with no options:

In [375]: df = pd.read_xml("https://www.w3schools.com/xml/books.xml")

In [376]: df
Out[376]: 
   category              title                  author  year  price      cover
0   cooking   Everyday Italian     Giada De Laurentiis  2005  30.00       None
1  children       Harry Potter            J K. Rowling  2005  29.99       None
2       web  XQuery Kick Start  Vaidyanathan Nagarajan  2003  49.99       None
3       web       Learning XML             Erik T. Ray  2003  39.95  paperback

Read in the content of the “books.xml” file and pass it to read_xml as a string:

In [377]: file_path = "books.xml"

In [378]: with open(file_path, "w") as f:
   .....:     f.write(xml)
   .....: 

In [379]: with open(file_path, "r") as f:
   .....:     df = pd.read_xml(StringIO(f.read()))
   .....: 

In [380]: df
Out[380]: 
   category             title               author  year  price
0   cooking  Everyday Italian  Giada De Laurentiis  2005  30.00
1  children      Harry Potter         J K. Rowling  2005  29.99
2       web      Learning XML          Erik T. Ray  2003  39.95

Read in the content of the “books.xml” as instance of StringIO or BytesIO and pass it to read_xml:

In [381]: with open(file_path, "r") as f:
   .....:     sio = StringIO(f.read())
   .....: 

In [382]: df = pd.read_xml(sio)

In [383]: df
Out[383]: 
   category             title               author  year  price
0   cooking  Everyday Italian  Giada De Laurentiis  2005  30.00
1  children      Harry Potter         J K. Rowling  2005  29.99
2       web      Learning XML          Erik T. Ray  2003  39.95
In [384]: with open(file_path, "rb") as f:
   .....:     bio = BytesIO(f.read())
   .....: 

In [385]: df = pd.read_xml(bio)

In [386]: df
Out[386]: 
   category             title               author  year  price
0   cooking  Everyday Italian  Giada De Laurentiis  2005  30.00
1  children      Harry Potter         J K. Rowling  2005  29.99
2       web      Learning XML          Erik T. Ray  2003  39.95

Even read XML from AWS S3 buckets such as NIH NCBI PMC Article Datasets providing Biomedical and Life Science Jorurnals:

In [387]: df = pd.read_xml(
   .....:     "s3://pmc-oa-opendata/oa_comm/xml/all/PMC1236943.xml",
   .....:     xpath=".//journal-meta",
   .....: )
   .....: 

In [388]: df
Out[388]: 
              journal-id              journal-title       issn  publisher
0  Cardiovasc Ultrasound  Cardiovascular Ultrasound  1476-7120        NaN

With lxml as default parser, you access the full-featured XML library that extends Python’s ElementTree API. One powerful tool is ability to query nodes selectively or conditionally with more expressive XPath:

In [389]: df = pd.read_xml(file_path, xpath="//book[year=2005]")

In [390]: df
Out[390]: 
   category             title               author  year  price
0   cooking  Everyday Italian  Giada De Laurentiis  2005  30.00
1  children      Harry Potter         J K. Rowling  2005  29.99

Specify only elements or only attributes to parse:

In [391]: df = pd.read_xml(file_path, elems_only=True)

In [392]: df
Out[392]: 
              title               author  year  price
0  Everyday Italian  Giada De Laurentiis  2005  30.00
1      Harry Potter         J K. Rowling  2005  29.99
2      Learning XML          Erik T. Ray  2003  39.95
In [393]: df = pd.read_xml(file_path, attrs_only=True)

In [394]: df
Out[394]: 
   category
0   cooking
1  children
2       web

XML documents can have namespaces with prefixes and default namespaces without prefixes both of which are denoted with a special attribute xmlns. In order to parse by node under a namespace context, xpath must reference a prefix.

For example, below XML contains a namespace with prefix, doc, and URI at https://example.com. In order to parse doc:row nodes, namespaces must be used.

In [395]: xml = """<?xml version='1.0' encoding='utf-8'?>
   .....: <doc:data xmlns:doc="https://example.com">
   .....:   <doc:row>
   .....:     <doc:shape>square</doc:shape>
   .....:     <doc:degrees>360</doc:degrees>
   .....:     <doc:sides>4.0</doc:sides>
   .....:   </doc:row>
   .....:   <doc:row>
   .....:     <doc:shape>circle</doc:shape>
   .....:     <doc:degrees>360</doc:degrees>
   .....:     <doc:sides/>
   .....:   </doc:row>
   .....:   <doc:row>
   .....:     <doc:shape>triangle</doc:shape>
   .....:     <doc:degrees>180</doc:degrees>
   .....:     <doc:sides>3.0</doc:sides>
   .....:   </doc:row>
   .....: </doc:data>"""
   .....: 

In [396]: df = pd.read_xml(StringIO(xml),
   .....:                  xpath="//doc:row",
   .....:                  namespaces={"doc": "https://example.com"})
   .....: 

In [397]: df
Out[397]: 
      shape  degrees  sides
0    square      360    4.0
1    circle      360    NaN
2  triangle      180    3.0

Similarly, an XML document can have a default namespace without prefix. Failing to assign a temporary prefix will return no nodes and raise a ValueError. But assigning any temporary name to correct URI allows parsing by nodes.

In [398]: xml = """<?xml version='1.0' encoding='utf-8'?>
   .....: <data xmlns="https://example.com">
   .....:  <row>
   .....:    <shape>square</shape>
   .....:    <degrees>360</degrees>
   .....:    <sides>4.0</sides>
   .....:  </row>
   .....:  <row>
   .....:    <shape>circle</shape>
   .....:    <degrees>360</degrees>
   .....:    <sides/>
   .....:  </row>
   .....:  <row>
   .....:    <shape>triangle</shape>
   .....:    <degrees>180</degrees>
   .....:    <sides>3.0</sides>
   .....:  </row>
   .....: </data>"""
   .....: 

In [399]: df = pd.read_xml(StringIO(xml),
   .....:                  xpath="//pandas:row",
   .....:                  namespaces={"pandas": "https://example.com"})
   .....: 

In [400]: df
Out[400]: 
      shape  degrees  sides
0    square      360    4.0
1    circle      360    NaN
2  triangle      180    3.0

However, if XPath does not reference node names such as default, /*, then namespaces is not required.

Note

Since xpath identifies the parent of content to be parsed, only immediate desendants which include child nodes or current attributes are parsed. Therefore, read_xml will not parse the text of grandchildren or other descendants and will not parse attributes of any descendant. To retrieve lower level content, adjust xpath to lower level. For example,

In [401]: xml = """
   .....: <data>
   .....:   <row>
   .....:     <shape sides="4">square</shape>
   .....:     <degrees>360</degrees>
   .....:   </row>
   .....:   <row>
   .....:     <shape sides="0">circle</shape>
   .....:     <degrees>360</degrees>
   .....:   </row>
   .....:   <row>
   .....:     <shape sides="3">triangle</shape>
   .....:     <degrees>180</degrees>
   .....:   </row>
   .....: </data>"""
   .....: 

In [402]: df = pd.read_xml(StringIO(xml), xpath="./row")

In [403]: df
Out[403]: 
      shape  degrees
0    square      360
1    circle      360
2  triangle      180

shows the attribute sides on shape element was not parsed as expected since this attribute resides on the child of row element and not row element itself. In other words, sides attribute is a grandchild level descendant of row element. However, the xpath targets row element which covers only its children and attributes.

With lxml as parser, you can flatten nested XML documents with an XSLT script which also can be string/file/URL types. As background, XSLT is a special-purpose language written in a special XML file that can transform original XML documents into other XML, HTML, even text (CSV, JSON, etc.) using an XSLT processor.

For example, consider this somewhat nested structure of Chicago “L” Rides where station and rides elements encapsulate data in their own sections. With below XSLT, lxml can transform original nested document into a flatter output (as shown below for demonstration) for easier parse into DataFrame:

In [404]: xml = """<?xml version='1.0' encoding='utf-8'?>
   .....:  <response>
   .....:   <row>
   .....:     <station id="40850" name="Library"/>
   .....:     <month>2020-09-01T00:00:00</month>
   .....:     <rides>
   .....:       <avg_weekday_rides>864.2</avg_weekday_rides>
   .....:       <avg_saturday_rides>534</avg_saturday_rides>
   .....:       <avg_sunday_holiday_rides>417.2</avg_sunday_holiday_rides>
   .....:     </rides>
   .....:   </row>
   .....:   <row>
   .....:     <station id="41700" name="Washington/Wabash"/>
   .....:     <month>2020-09-01T00:00:00</month>
   .....:     <rides>
   .....:       <avg_weekday_rides>2707.4</avg_weekday_rides>
   .....:       <avg_saturday_rides>1909.8</avg_saturday_rides>
   .....:       <avg_sunday_holiday_rides>1438.6</avg_sunday_holiday_rides>
   .....:     </rides>
   .....:   </row>
   .....:   <row>
   .....:     <station id="40380" name="Clark/Lake"/>
   .....:     <month>2020-09-01T00:00:00</month>
   .....:     <rides>
   .....:       <avg_weekday_rides>2949.6</avg_weekday_rides>
   .....:       <avg_saturday_rides>1657</avg_saturday_rides>
   .....:       <avg_sunday_holiday_rides>1453.8</avg_sunday_holiday_rides>
   .....:     </rides>
   .....:   </row>
   .....:  </response>"""
   .....: 

In [405]: xsl = """<xsl:stylesheet version="1.0" xmlns:xsl="http://www.w3.org/1999/XSL/Transform">
   .....:    <xsl:output method="xml" omit-xml-declaration="no" indent="yes"/>
   .....:    <xsl:strip-space elements="*"/>
   .....:    <xsl:template match="/response">
   .....:       <xsl:copy>
   .....:         <xsl:apply-templates select="row"/>
   .....:       </xsl:copy>
   .....:    </xsl:template>
   .....:    <xsl:template match="row">
   .....:       <xsl:copy>
   .....:         <station_id><xsl:value-of select="station/@id"/></station_id>
   .....:         <station_name><xsl:value-of select="station/@name"/></station_name>
   .....:         <xsl:copy-of select="month|rides/*"/>
   .....:       </xsl:copy>
   .....:    </xsl:template>
   .....:  </xsl:stylesheet>"""
   .....: 

In [406]: output = """<?xml version='1.0' encoding='utf-8'?>
   .....:  <response>
   .....:    <row>
   .....:       <station_id>40850</station_id>
   .....:       <station_name>Library</station_name>
   .....:       <month>2020-09-01T00:00:00</month>
   .....:       <avg_weekday_rides>864.2</avg_weekday_rides>
   .....:       <avg_saturday_rides>534</avg_saturday_rides>
   .....:       <avg_sunday_holiday_rides>417.2</avg_sunday_holiday_rides>
   .....:    </row>
   .....:    <row>
   .....:       <station_id>41700</station_id>
   .....:       <station_name>Washington/Wabash</station_name>
   .....:       <month>2020-09-01T00:00:00</month>
   .....:       <avg_weekday_rides>2707.4</avg_weekday_rides>
   .....:       <avg_saturday_rides>1909.8</avg_saturday_rides>
   .....:       <avg_sunday_holiday_rides>1438.6</avg_sunday_holiday_rides>
   .....:    </row>
   .....:    <row>
   .....:       <station_id>40380</station_id>
   .....:       <station_name>Clark/Lake</station_name>
   .....:       <month>2020-09-01T00:00:00</month>
   .....:       <avg_weekday_rides>2949.6</avg_weekday_rides>
   .....:       <avg_saturday_rides>1657</avg_saturday_rides>
   .....:       <avg_sunday_holiday_rides>1453.8</avg_sunday_holiday_rides>
   .....:    </row>
   .....:  </response>"""
   .....: 

In [407]: df = pd.read_xml(StringIO(xml), stylesheet=xsl)

In [408]: df
Out[408]: 
   station_id       station_name  ... avg_saturday_rides  avg_sunday_holiday_rides
0       40850            Library  ...              534.0                     417.2
1       41700  Washington/Wabash  ...             1909.8                    1438.6
2       40380         Clark/Lake  ...             1657.0                    1453.8

[3 rows x 6 columns]

For very large XML files that can range in hundreds of megabytes to gigabytes, pandas.read_xml() supports parsing such sizeable files using lxml’s iterparse and etree’s iterparse which are memory-efficient methods to iterate through an XML tree and extract specific elements and attributes. without holding entire tree in memory.

New in version 1.5.0.

To use this feature, you must pass a physical XML file path into read_xml and use the iterparse argument. Files should not be compressed or point to online sources but stored on local disk. Also, iterparse should be a dictionary where the key is the repeating nodes in document (which become the rows) and the value is a list of any element or attribute that is a descendant (i.e., child, grandchild) of repeating node. Since XPath is not used in this method, descendants do not need to share same relationship with one another. Below shows example of reading in Wikipedia’s very large (12 GB+) latest article data dump.

In [1]: df = pd.read_xml(
...         "/path/to/downloaded/enwikisource-latest-pages-articles.xml",
...         iterparse = {"page": ["title", "ns", "id"]}
...     )
...     df
Out[2]:
                                                     title   ns        id
0                                       Gettysburg Address    0     21450
1                                                Main Page    0     42950
2                            Declaration by United Nations    0      8435
3             Constitution of the United States of America    0      8435
4                     Declaration of Independence (Israel)    0     17858
...                                                    ...  ...       ...
3578760               Page:Black cat 1897 07 v2 n10.pdf/17  104    219649
3578761               Page:Black cat 1897 07 v2 n10.pdf/43  104    219649
3578762               Page:Black cat 1897 07 v2 n10.pdf/44  104    219649
3578763      The History of Tom Jones, a Foundling/Book IX    0  12084291
3578764  Page:Shakespeare of Stratford (1926) Yale.djvu/91  104     21450

[3578765 rows x 3 columns]

Writing XML#

New in version 1.3.0.

DataFrame objects have an instance method to_xml which renders the contents of the DataFrame as an XML document.

Note

This method does not support special properties of XML including DTD, CData, XSD schemas, processing instructions, comments, and others. Only namespaces at the root level is supported. However, stylesheet allows design changes after initial output.

Let’s look at a few examples.

Write an XML without options:

In [409]: geom_df = pd.DataFrame(
   .....:     {
   .....:         "shape": ["square", "circle", "triangle"],
   .....:         "degrees": [360, 360, 180],
   .....:         "sides": [4, np.nan, 3],
   .....:     }
   .....: )
   .....: 

In [410]: print(geom_df.to_xml())
<?xml version='1.0' encoding='utf-8'?>
<data>
  <row>
    <index>0</index>
    <shape>square</shape>
    <degrees>360</degrees>
    <sides>4.0</sides>
  </row>
  <row>
    <index>1</index>
    <shape>circle</shape>
    <degrees>360</degrees>
    <sides/>
  </row>
  <row>
    <index>2</index>
    <shape>triangle</shape>
    <degrees>180</degrees>
    <sides>3.0</sides>
  </row>
</data>

Write an XML with new root and row name:

In [411]: print(geom_df.to_xml(root_name="geometry", row_name="objects"))
<?xml version='1.0' encoding='utf-8'?>
<geometry>
  <objects>
    <index>0</index>
    <shape>square</shape>
    <degrees>360</degrees>
    <sides>4.0</sides>
  </objects>
  <objects>
    <index>1</index>
    <shape>circle</shape>
    <degrees>360</degrees>
    <sides/>
  </objects>
  <objects>
    <index>2</index>
    <shape>triangle</shape>
    <degrees>180</degrees>
    <sides>3.0</sides>
  </objects>
</geometry>

Write an attribute-centric XML:

In [412]: print(geom_df.to_xml(attr_cols=geom_df.columns.tolist()))
<?xml version='1.0' encoding='utf-8'?>
<data>
  <row index="0" shape="square" degrees="360" sides="4.0"/>
  <row index="1" shape="circle" degrees="360"/>
  <row index="2" shape="triangle" degrees="180" sides="3.0"/>
</data>

Write a mix of elements and attributes:

In [413]: print(
   .....:     geom_df.to_xml(
   .....:         index=False,
   .....:         attr_cols=['shape'],
   .....:         elem_cols=['degrees', 'sides'])
   .....: )
   .....: 
<?xml version='1.0' encoding='utf-8'?>
<data>
  <row shape="square">
    <degrees>360</degrees>
    <sides>4.0</sides>
  </row>
  <row shape="circle">
    <degrees>360</degrees>
    <sides/>
  </row>
  <row shape="triangle">
    <degrees>180</degrees>
    <sides>3.0</sides>
  </row>
</data>

Any DataFrames with hierarchical columns will be flattened for XML element names with levels delimited by underscores:

In [414]: ext_geom_df = pd.DataFrame(
   .....:     {
   .....:         "type": ["polygon", "other", "polygon"],
   .....:         "shape": ["square", "circle", "triangle"],
   .....:         "degrees": [360, 360, 180],
   .....:         "sides": [4, np.nan, 3],
   .....:     }
   .....: )
   .....: 

In [415]: pvt_df = ext_geom_df.pivot_table(index='shape',
   .....:                                  columns='type',
   .....:                                  values=['degrees', 'sides'],
   .....:                                  aggfunc='sum')
   .....: 

In [416]: pvt_df
Out[416]: 
         degrees         sides        
type       other polygon other polygon
shape                                 
circle     360.0     NaN   0.0     NaN
square       NaN   360.0   NaN     4.0
triangle     NaN   180.0   NaN     3.0

In [417]: print(pvt_df.to_xml())
<?xml version='1.0' encoding='utf-8'?>
<data>
  <row>
    <shape>circle</shape>
    <degrees_other>360.0</degrees_other>
    <degrees_polygon/>
    <sides_other>0.0</sides_other>
    <sides_polygon/>
  </row>
  <row>
    <shape>square</shape>
    <degrees_other/>
    <degrees_polygon>360.0</degrees_polygon>
    <sides_other/>
    <sides_polygon>4.0</sides_polygon>
  </row>
  <row>
    <shape>triangle</shape>
    <degrees_other/>
    <degrees_polygon>180.0</degrees_polygon>
    <sides_other/>
    <sides_polygon>3.0</sides_polygon>
  </row>
</data>

Write an XML with default namespace:

In [418]: print(geom_df.to_xml(namespaces={"": "https://example.com"}))
<?xml version='1.0' encoding='utf-8'?>
<data xmlns="https://example.com">
  <row>
    <index>0</index>
    <shape>square</shape>
    <degrees>360</degrees>
    <sides>4.0</sides>
  </row>
  <row>
    <index>1</index>
    <shape>circle</shape>
    <degrees>360</degrees>
    <sides/>
  </row>
  <row>
    <index>2</index>
    <shape>triangle</shape>
    <degrees>180</degrees>
    <sides>3.0</sides>
  </row>
</data>

Write an XML with namespace prefix:

In [419]: print(
   .....:     geom_df.to_xml(namespaces={"doc": "https://example.com"},
   .....:                    prefix="doc")
   .....: )
   .....: 
<?xml version='1.0' encoding='utf-8'?>
<doc:data xmlns:doc="https://example.com">
  <doc:row>
    <doc:index>0</doc:index>
    <doc:shape>square</doc:shape>
    <doc:degrees>360</doc:degrees>
    <doc:sides>4.0</doc:sides>
  </doc:row>
  <doc:row>
    <doc:index>1</doc:index>
    <doc:shape>circle</doc:shape>
    <doc:degrees>360</doc:degrees>
    <doc:sides/>
  </doc:row>
  <doc:row>
    <doc:index>2</doc:index>
    <doc:shape>triangle</doc:shape>
    <doc:degrees>180</doc:degrees>
    <doc:sides>3.0</doc:sides>
  </doc:row>
</doc:data>

Write an XML without declaration or pretty print:

In [420]: print(
   .....:     geom_df.to_xml(xml_declaration=False,
   .....:                    pretty_print=False)
   .....: )
   .....: 
<data><row><index>0</index><shape>square</shape><degrees>360</degrees><sides>4.0</sides></row><row><index>1</index><shape>circle</shape><degrees>360</degrees><sides/></row><row><index>2</index><shape>triangle</shape><degrees>180</degrees><sides>3.0</sides></row></data>

Write an XML and transform with stylesheet:

In [421]: xsl = """<xsl:stylesheet version="1.0" xmlns:xsl="http://www.w3.org/1999/XSL/Transform">
   .....:    <xsl:output method="xml" omit-xml-declaration="no" indent="yes"/>
   .....:    <xsl:strip-space elements="*"/>
   .....:    <xsl:template match="/data">
   .....:      <geometry>
   .....:        <xsl:apply-templates select="row"/>
   .....:      </geometry>
   .....:    </xsl:template>
   .....:    <xsl:template match="row">
   .....:      <object index="{index}">
   .....:        <xsl:if test="shape!='circle'">
   .....:            <xsl:attribute name="type">polygon</xsl:attribute>
   .....:        </xsl:if>
   .....:        <xsl:copy-of select="shape"/>
   .....:        <property>
   .....:            <xsl:copy-of select="degrees|sides"/>
   .....:        </property>
   .....:      </object>
   .....:    </xsl:template>
   .....:  </xsl:stylesheet>"""
   .....: 

In [422]: print(geom_df.to_xml(stylesheet=xsl))
<?xml version="1.0"?>
<geometry>
  <object index="0" type="polygon">
    <shape>square</shape>
    <property>
      <degrees>360</degrees>
      <sides>4.0</sides>
    </property>
  </object>
  <object index="1">
    <shape>circle</shape>
    <property>
      <degrees>360</degrees>
      <sides/>
    </property>
  </object>
  <object index="2" type="polygon">
    <shape>triangle</shape>
    <property>
      <degrees>180</degrees>
      <sides>3.0</sides>
    </property>
  </object>
</geometry>

XML Final Notes#

  • All XML documents adhere to W3C specifications. Both etree and lxml parsers will fail to parse any markup document that is not well-formed or follows XML syntax rules. Do be aware HTML is not an XML document unless it follows XHTML specs. However, other popular markup types including KML, XAML, RSS, MusicML, MathML are compliant XML schemas.

  • For above reason, if your application builds XML prior to pandas operations, use appropriate DOM libraries like etree and lxml to build the necessary document and not by string concatenation or regex adjustments. Always remember XML is a special text file with markup rules.

  • With very large XML files (several hundred MBs to GBs), XPath and XSLT can become memory-intensive operations. Be sure to have enough available RAM for reading and writing to large XML files (roughly about 5 times the size of text).

  • Because XSLT is a programming language, use it with caution since such scripts can pose a security risk in your environment and can run large or infinite recursive operations. Always test scripts on small fragments before full run.

  • The etree parser supports all functionality of both read_xml and to_xml except for complex XPath and any XSLT. Though limited in features, etree is still a reliable and capable parser and tree builder. Its performance may trail lxml to a certain degree for larger files but relatively unnoticeable on small to medium size files.

Excel files#

The read_excel() method can read Excel 2007+ (.xlsx) files using the openpyxl Python module. Excel 2003 (.xls) files can be read using xlrd. Binary Excel (.xlsb) files can be read using pyxlsb. The to_excel() instance method is used for saving a DataFrame to Excel. Generally the semantics are similar to working with csv data. See the cookbook for some advanced strategies.

Warning

The xlrd package is now only for reading old-style .xls files.

Before pandas 1.3.0, the default argument engine=None to read_excel() would result in using the xlrd engine in many cases, including new Excel 2007+ (.xlsx) files. pandas will now default to using the openpyxl engine.

It is strongly encouraged to install openpyxl to read Excel 2007+ (.xlsx) files. Please do not report issues when using ``xlrd`` to read ``.xlsx`` files. This is no longer supported, switch to using openpyxl instead.

Reading Excel files#

In the most basic use-case, read_excel takes a path to an Excel file, and the sheet_name indicating which sheet to parse.

When using the engine_kwargs parameter, pandas will pass these arguments to the engine. For this, it is important to know which function pandas is using internally.

  • For the engine openpyxl, pandas is using openpyxl.load_workbook() to read in (.xlsx) and (.xlsm) files.

  • For the engine xlrd, pandas is using xlrd.open_workbook() to read in (.xls) files.

  • For the engine pyxlsb, pandas is using pyxlsb.open_workbook() to read in (.xlsb) files.

  • For the engine odf, pandas is using odf.opendocument.load() to read in (.ods) files.

# Returns a DataFrame
pd.read_excel("path_to_file.xls", sheet_name="Sheet1")

ExcelFile class#

To facilitate working with multiple sheets from the same file, the ExcelFile class can be used to wrap the file and can be passed into read_excel There will be a performance benefit for reading multiple sheets as the file is read into memory only once.

xlsx = pd.ExcelFile("path_to_file.xls")
df = pd.read_excel(xlsx, "Sheet1")

The ExcelFile class can also be used as a context manager.

with pd.ExcelFile("path_to_file.xls") as xls:
    df1 = pd.read_excel(xls, "Sheet1")
    df2 = pd.read_excel(xls, "Sheet2")

The sheet_names property will generate a list of the sheet names in the file.

The primary use-case for an ExcelFile is parsing multiple sheets with different parameters:

data = {}
# For when Sheet1's format differs from Sheet2
with pd.ExcelFile("path_to_file.xls") as xls:
    data["Sheet1"] = pd.read_excel(xls, "Sheet1", index_col=None, na_values=["NA"])
    data["Sheet2"] = pd.read_excel(xls, "Sheet2", index_col=1)

Note that if the same parsing parameters are used for all sheets, a list of sheet names can simply be passed to read_excel with no loss in performance.

# using the ExcelFile class
data = {}
with pd.ExcelFile("path_to_file.xls") as xls:
    data["Sheet1"] = pd.read_excel(xls, "Sheet1", index_col=None, na_values=["NA"])
    data["Sheet2"] = pd.read_excel(xls, "Sheet2", index_col=None, na_values=["NA"])

# equivalent using the read_excel function
data = pd.read_excel(
    "path_to_file.xls", ["Sheet1", "Sheet2"], index_col=None, na_values=["NA"]
)

ExcelFile can also be called with a xlrd.book.Book object as a parameter. This allows the user to control how the excel file is read. For example, sheets can be loaded on demand by calling xlrd.open_workbook() with on_demand=True.

import xlrd

xlrd_book = xlrd.open_workbook("path_to_file.xls", on_demand=True)
with pd.ExcelFile(xlrd_book) as xls:
    df1 = pd.read_excel(xls, "Sheet1")
    df2 = pd.read_excel(xls, "Sheet2")

Specifying sheets#

Note

The second argument is sheet_name, not to be confused with ExcelFile.sheet_names.

Note

An ExcelFile’s attribute sheet_names provides access to a list of sheets.

  • The arguments sheet_name allows specifying the sheet or sheets to read.

  • The default value for sheet_name is 0, indicating to read the first sheet

  • Pass a string to refer to the name of a particular sheet in the workbook.

  • Pass an integer to refer to the index of a sheet. Indices follow Python convention, beginning at 0.

  • Pass a list of either strings or integers, to return a dictionary of specified sheets.

  • Pass a None to return a dictionary of all available sheets.

# Returns a DataFrame
pd.read_excel("path_to_file.xls", "Sheet1", index_col=None, na_values=["NA"])

Using the sheet index:

# Returns a DataFrame
pd.read_excel("path_to_file.xls", 0, index_col=None, na_values=["NA"])

Using all default values:

# Returns a DataFrame
pd.read_excel("path_to_file.xls")

Using None to get all sheets:

# Returns a dictionary of DataFrames
pd.read_excel("path_to_file.xls", sheet_name=None)

Using a list to get multiple sheets:

# Returns the 1st and 4th sheet, as a dictionary of DataFrames.
pd.read_excel("path_to_file.xls", sheet_name=["Sheet1", 3])

read_excel can read more than one sheet, by setting sheet_name to either a list of sheet names, a list of sheet positions, or None to read all sheets. Sheets can be specified by sheet index or sheet name, using an integer or string, respectively.

Reading a MultiIndex#

read_excel can read a MultiIndex index, by passing a list of columns to index_col and a MultiIndex column by passing a list of rows to header. If either the index or columns have serialized level names those will be read in as well by specifying the rows/columns that make up the levels.

For example, to read in a MultiIndex index without names:

In [423]: df = pd.DataFrame(
   .....:     {"a": [1, 2, 3, 4], "b": [5, 6, 7, 8]},
   .....:     index=pd.MultiIndex.from_product([["a", "b"], ["c", "d"]]),
   .....: )
   .....: 

In [424]: df.to_excel("path_to_file.xlsx")

In [425]: df = pd.read_excel("path_to_file.xlsx", index_col=[0, 1])

In [426]: df
Out[426]: 
     a  b
a c  1  5
  d  2  6
b c  3  7
  d  4  8

If the index has level names, they will parsed as well, using the same parameters.

In [427]: df.index = df.index.set_names(["lvl1", "lvl2"])

In [428]: df.to_excel("path_to_file.xlsx")

In [429]: df = pd.read_excel("path_to_file.xlsx", index_col=[0, 1])

In [430]: df
Out[430]: 
           a  b
lvl1 lvl2      
a    c     1  5
     d     2  6
b    c     3  7
     d     4  8

If the source file has both MultiIndex index and columns, lists specifying each should be passed to index_col and header:

In [431]: df.columns = pd.MultiIndex.from_product([["a"], ["b", "d"]], names=["c1", "c2"])

In [432]: df.to_excel("path_to_file.xlsx")

In [433]: df = pd.read_excel("path_to_file.xlsx", index_col=[0, 1], header=[0, 1])

In [434]: df
Out[434]: 
c1         a   
c2         b  d
lvl1 lvl2      
a    c     1  5
     d     2  6
b    c     3  7
     d     4  8

Missing values in columns specified in index_col will be forward filled to allow roundtripping with to_excel for merged_cells=True. To avoid forward filling the missing values use set_index after reading the data instead of index_col.

Parsing specific columns#

It is often the case that users will insert columns to do temporary computations in Excel and you may not want to read in those columns. read_excel takes a usecols keyword to allow you to specify a subset of columns to parse.

You can specify a comma-delimited set of Excel columns and ranges as a string:

pd.read_excel("path_to_file.xls", "Sheet1", usecols="A,C:E")

If usecols is a list of integers, then it is assumed to be the file column indices to be parsed.

pd.read_excel("path_to_file.xls", "Sheet1", usecols=[0, 2, 3])

Element order is ignored, so usecols=[0, 1] is the same as [1, 0].

If usecols is a list of strings, it is assumed that each string corresponds to a column name provided either by the user in names or inferred from the document header row(s). Those strings define which columns will be parsed:

pd.read_excel("path_to_file.xls", "Sheet1", usecols=["foo", "bar"])

Element order is ignored, so usecols=['baz', 'joe'] is the same as ['joe', 'baz'].

If usecols is callable, the callable function will be evaluated against the column names, returning names where the callable function evaluates to True.

pd.read_excel("path_to_file.xls", "Sheet1", usecols=lambda x: x.isalpha())

Parsing dates#

Datetime-like values are normally automatically converted to the appropriate dtype when reading the excel file. But if you have a column of strings that look like dates (but are not actually formatted as dates in excel), you can use the parse_dates keyword to parse those strings to datetimes:

pd.read_excel("path_to_file.xls", "Sheet1", parse_dates=["date_strings"])

Cell converters#

It is possible to transform the contents of Excel cells via the converters option. For instance, to convert a column to boolean:

pd.read_excel("path_to_file.xls", "Sheet1", converters={"MyBools": bool})

This options handles missing values and treats exceptions in the converters as missing data. Transformations are applied cell by cell rather than to the column as a whole, so the array dtype is not guaranteed. For instance, a column of integers with missing values cannot be transformed to an array with integer dtype, because NaN is strictly a float. You can manually mask missing data to recover integer dtype:

def cfun(x):
    return int(x) if x else -1


pd.read_excel("path_to_file.xls", "Sheet1", converters={"MyInts": cfun})

Dtype specifications#

As an alternative to converters, the type for an entire column can be specified using the dtype keyword, which takes a dictionary mapping column names to types. To interpret data with no type inference, use the type str or object.

pd.read_excel("path_to_file.xls", dtype={"MyInts": "int64", "MyText": str})

Writing Excel files#

Writing Excel files to disk#

To write a DataFrame object to a sheet of an Excel file, you can use the to_excel instance method. The arguments are largely the same as to_csv described above, the first argument being the name of the excel file, and the optional second argument the name of the sheet to which the DataFrame should be written. For example:

df.to_excel("path_to_file.xlsx", sheet_name="Sheet1")

Files with a .xlsx extension will be written using xlsxwriter (if available) or openpyxl.

The DataFrame will be written in a way that tries to mimic the REPL output. The index_label will be placed in the second row instead of the first. You can place it in the first row by setting the merge_cells option in to_excel() to False:

df.to_excel("path_to_file.xlsx", index_label="label", merge_cells=False)

In order to write separate DataFrames to separate sheets in a single Excel file, one can pass an ExcelWriter.

with pd.ExcelWriter("path_to_file.xlsx") as writer:
    df1.to_excel(writer, sheet_name="Sheet1")
    df2.to_excel(writer, sheet_name="Sheet2")

When using the engine_kwargs parameter, pandas will pass these arguments to the engine. For this, it is important to know which function pandas is using internally.

  • For the engine openpyxl, pandas is using openpyxl.Workbook() to create a new sheet and openpyxl.load_workbook() to append data to an existing sheet. The openpyxl engine writes to (.xlsx) and (.xlsm) files.

  • For the engine xlsxwriter, pandas is using xlsxwriter.Workbook() to write to (.xlsx) files.

  • For the engine odf, pandas is using odf.opendocument.OpenDocumentSpreadsheet() to write to (.ods) files.

Writing Excel files to memory#

pandas supports writing Excel files to buffer-like objects such as StringIO or BytesIO using ExcelWriter.

from io import BytesIO

bio = BytesIO()

# By setting the 'engine' in the ExcelWriter constructor.
writer = pd.ExcelWriter(bio, engine="xlsxwriter")
df.to_excel(writer, sheet_name="Sheet1")

# Save the workbook
writer.save()

# Seek to the beginning and read to copy the workbook to a variable in memory
bio.seek(0)
workbook = bio.read()

Note

engine is optional but recommended. Setting the engine determines the version of workbook produced. Setting engine='xlrd' will produce an Excel 2003-format workbook (xls). Using either 'openpyxl' or 'xlsxwriter' will produce an Excel 2007-format workbook (xlsx). If omitted, an Excel 2007-formatted workbook is produced.

Excel writer engines#

pandas chooses an Excel writer via two methods:

  1. the engine keyword argument

  2. the filename extension (via the default specified in config options)

By default, pandas uses the XlsxWriter for .xlsx, openpyxl for .xlsm. If you have multiple engines installed, you can set the default engine through setting the config options io.excel.xlsx.writer and io.excel.xls.writer. pandas will fall back on openpyxl for .xlsx files if Xlsxwriter is not available.

To specify which writer you want to use, you can pass an engine keyword argument to to_excel and to ExcelWriter. The built-in engines are:

  • openpyxl: version 2.4 or higher is required

  • xlsxwriter

# By setting the 'engine' in the DataFrame 'to_excel()' methods.
df.to_excel("path_to_file.xlsx", sheet_name="Sheet1", engine="xlsxwriter")

# By setting the 'engine' in the ExcelWriter constructor.
writer = pd.ExcelWriter("path_to_file.xlsx", engine="xlsxwriter")

# Or via pandas configuration.
from pandas import options  # noqa: E402

options.io.excel.xlsx.writer = "xlsxwriter"

df.to_excel("path_to_file.xlsx", sheet_name="Sheet1")

Style and formatting#

The look and feel of Excel worksheets created from pandas can be modified using the following parameters on the DataFrame’s to_excel method.

  • float_format : Format string for floating point numbers (default None).

  • freeze_panes : A tuple of two integers representing the bottommost row and rightmost column to freeze. Each of these parameters is one-based, so (1, 1) will freeze the first row and first column (default None).

Using the Xlsxwriter engine provides many options for controlling the format of an Excel worksheet created with the to_excel method. Excellent examples can be found in the Xlsxwriter documentation here: https://xlsxwriter.readthedocs.io/working_with_pandas.html

OpenDocument Spreadsheets#

The io methods for Excel files also support reading and writing OpenDocument spreadsheets using the odfpy module. The semantics and features for reading and writing OpenDocument spreadsheets match what can be done for Excel files using engine='odf'. The optional dependency ‘odfpy’ needs to be installed.

The read_excel() method can read OpenDocument spreadsheets

# Returns a DataFrame
pd.read_excel("path_to_file.ods", engine="odf")

Similarly, the to_excel() method can write OpenDocument spreadsheets

# Writes DataFrame to a .ods file
df.to_excel("path_to_file.ods", engine="odf")

Binary Excel (.xlsb) files#

The read_excel() method can also read binary Excel files using the pyxlsb module. The semantics and features for reading binary Excel files mostly match what can be done for Excel files using engine='pyxlsb'. pyxlsb does not recognize datetime types in files and will return floats instead.

# Returns a DataFrame
pd.read_excel("path_to_file.xlsb", engine="pyxlsb")

Note

Currently pandas only supports reading binary Excel files. Writing is not implemented.

Clipboard#

A handy way to grab data is to use the read_clipboard() method, which takes the contents of the clipboard buffer and passes them to the read_csv method. For instance, you can copy the following text to the clipboard (CTRL-C on many operating systems):

  A B C
x 1 4 p
y 2 5 q
z 3 6 r

And then import the data directly to a DataFrame by calling:

>>> clipdf = pd.read_clipboard()
>>> clipdf
  A B C
x 1 4 p
y 2 5 q
z 3 6 r

The to_clipboard method can be used to write the contents of a DataFrame to the clipboard. Following which you can paste the clipboard contents into other applications (CTRL-V on many operating systems). Here we illustrate writing a DataFrame into clipboard and reading it back.

>>> df = pd.DataFrame(
...     {"A": [1, 2, 3], "B": [4, 5, 6], "C": ["p", "q", "r"]}, index=["x", "y", "z"]
... )

>>> df
  A B C
x 1 4 p
y 2 5 q
z 3 6 r
>>> df.to_clipboard()
>>> pd.read_clipboard()
  A B C
x 1 4 p
y 2 5 q
z 3 6 r

We can see that we got the same content back, which we had earlier written to the clipboard.

Note

You may need to install xclip or xsel (with PyQt5, PyQt4 or qtpy) on Linux to use these methods.

Pickling#

All pandas objects are equipped with to_pickle methods which use Python’s cPickle module to save data structures to disk using the pickle format.

In [435]: df
Out[435]: 
c1         a   
c2         b  d
lvl1 lvl2      
a    c     1  5
     d     2  6
b    c     3  7
     d     4  8

In [436]: df.to_pickle("foo.pkl")

The read_pickle function in the pandas namespace can be used to load any pickled pandas object (or any other pickled object) from file:

In [437]: pd.read_pickle("foo.pkl")
Out[437]: 
c1         a   
c2         b  d
lvl1 lvl2      
a    c     1  5
     d     2  6
b    c     3  7
     d     4  8

Warning

Loading pickled data received from untrusted sources can be unsafe.

See: https://docs.python.org/3/library/pickle.html

Warning

read_pickle() is only guaranteed backwards compatible back to a few minor release.

Compressed pickle files#

read_pickle(), DataFrame.to_pickle() and Series.to_pickle() can read and write compressed pickle files. The compression types of gzip, bz2, xz, zstd are supported for reading and writing. The zip file format only supports reading and must contain only one data file to be read.

The compression type can be an explicit parameter or be inferred from the file extension. If ‘infer’, then use gzip, bz2, zip, xz, zstd if filename ends in '.gz', '.bz2', '.zip', '.xz', or '.zst', respectively.

The compression parameter can also be a dict in order to pass options to the compression protocol. It must have a 'method' key set to the name of the compression protocol, which must be one of {'zip', 'gzip', 'bz2', 'xz', 'zstd'}. All other key-value pairs are passed to the underlying compression library.

In [438]: df = pd.DataFrame(
   .....:     {
   .....:         "A": np.random.randn(1000),
   .....:         "B": "foo",
   .....:         "C": pd.date_range("20130101", periods=1000, freq="s"),
   .....:     }
   .....: )
   .....: 

In [439]: df
Out[439]: 
            A    B                   C
0   -0.317441  foo 2013-01-01 00:00:00
1   -1.236269  foo 2013-01-01 00:00:01
2    0.896171  foo 2013-01-01 00:00:02
3   -0.487602  foo 2013-01-01 00:00:03
4   -0.082240  foo 2013-01-01 00:00:04
..        ...  ...                 ...
995 -0.171092  foo 2013-01-01 00:16:35
996  1.786173  foo 2013-01-01 00:16:36
997 -0.575189  foo 2013-01-01 00:16:37
998  0.820750  foo 2013-01-01 00:16:38
999 -1.256530  foo 2013-01-01 00:16:39

[1000 rows x 3 columns]

Using an explicit compression type:

In [440]: df.to_pickle("data.pkl.compress", compression="gzip")

In [441]: rt = pd.read_pickle("data.pkl.compress", compression="gzip")

In [442]: rt
Out[442]: 
            A    B                   C
0   -0.317441  foo 2013-01-01 00:00:00
1   -1.236269  foo 2013-01-01 00:00:01
2    0.896171  foo 2013-01-01 00:00:02
3   -0.487602  foo 2013-01-01 00:00:03
4   -0.082240  foo 2013-01-01 00:00:04
..        ...  ...                 ...
995 -0.171092  foo 2013-01-01 00:16:35
996  1.786173  foo 2013-01-01 00:16:36
997 -0.575189  foo 2013-01-01 00:16:37
998  0.820750  foo 2013-01-01 00:16:38
999 -1.256530  foo 2013-01-01 00:16:39

[1000 rows x 3 columns]

Inferring compression type from the extension:

In [443]: df.to_pickle("data.pkl.xz", compression="infer")

In [444]: rt = pd.read_pickle("data.pkl.xz", compression="infer")

In [445]: rt
Out[445]: 
            A    B                   C
0   -0.317441  foo 2013-01-01 00:00:00
1   -1.236269  foo 2013-01-01 00:00:01
2    0.896171  foo 2013-01-01 00:00:02
3   -0.487602  foo 2013-01-01 00:00:03
4   -0.082240  foo 2013-01-01 00:00:04
..        ...  ...                 ...
995 -0.171092  foo 2013-01-01 00:16:35
996  1.786173  foo 2013-01-01 00:16:36
997 -0.575189  foo 2013-01-01 00:16:37
998  0.820750  foo 2013-01-01 00:16:38
999 -1.256530  foo 2013-01-01 00:16:39

[1000 rows x 3 columns]

The default is to ‘infer’:

In [446]: df.to_pickle("data.pkl.gz")

In [447]: rt = pd.read_pickle("data.pkl.gz")

In [448]: rt
Out[448]: 
            A    B                   C
0   -0.317441  foo 2013-01-01 00:00:00
1   -1.236269  foo 2013-01-01 00:00:01
2    0.896171  foo 2013-01-01 00:00:02
3   -0.487602  foo 2013-01-01 00:00:03
4   -0.082240  foo 2013-01-01 00:00:04
..        ...  ...                 ...
995 -0.171092  foo 2013-01-01 00:16:35
996  1.786173  foo 2013-01-01 00:16:36
997 -0.575189  foo 2013-01-01 00:16:37
998  0.820750  foo 2013-01-01 00:16:38
999 -1.256530  foo 2013-01-01 00:16:39

[1000 rows x 3 columns]

In [449]: df["A"].to_pickle("s1.pkl.bz2")

In [450]: rt = pd.read_pickle("s1.pkl.bz2")

In [451]: rt
Out[451]: 
0     -0.317441
1     -1.236269
2      0.896171
3     -0.487602
4     -0.082240
         ...   
995   -0.171092
996    1.786173
997   -0.575189
998    0.820750
999   -1.256530
Name: A, Length: 1000, dtype: float64

Passing options to the compression protocol in order to speed up compression:

In [452]: df.to_pickle("data.pkl.gz", compression={"method": "gzip", "compresslevel": 1})

msgpack#

pandas support for msgpack has been removed in version 1.0.0. It is recommended to use pickle instead.

Alternatively, you can also the Arrow IPC serialization format for on-the-wire transmission of pandas objects. For documentation on pyarrow, see here.

HDF5 (PyTables)#

HDFStore is a dict-like object which reads and writes pandas using the high performance HDF5 format using the excellent PyTables library. See the cookbook for some advanced strategies

Warning

pandas uses PyTables for reading and writing HDF5 files, which allows serializing object-dtype data with pickle. Loading pickled data received from untrusted sources can be unsafe.

See: https://docs.python.org/3/library/pickle.html for more.

In [453]: store = pd.HDFStore("store.h5")

In [454]: print(store)
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

Objects can be written to the file just like adding key-value pairs to a dict:

In [455]: index = pd.date_range("1/1/2000", periods=8)

In [456]: s = pd.Series(np.random.randn(5), index=["a", "b", "c", "d", "e"])

In [457]: df = pd.DataFrame(np.random.randn(8, 3), index=index, columns=["A", "B", "C"])

# store.put('s', s) is an equivalent method
In [458]: store["s"] = s

In [459]: store["df"] = df

In [460]: store
Out[460]: 
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

In a current or later Python session, you can retrieve stored objects:

# store.get('df') is an equivalent method
In [461]: store["df"]
Out[461]: 
                   A         B         C
2000-01-01  0.858644 -0.851236  1.058006
2000-01-02 -0.080372 -1.268121  1.561967
2000-01-03  0.816983  1.965656 -1.169408
2000-01-04  0.712795 -0.062433  0.736755
2000-01-05 -0.298721 -1.988045  1.475308
2000-01-06  1.103675  1.382242 -0.650762
2000-01-07 -0.729161 -0.142928 -1.063038
2000-01-08 -1.005977  0.465222 -0.094517

# dotted (attribute) access provides get as well
In [462]: store.df
Out[462]: 
                   A         B         C
2000-01-01  0.858644 -0.851236  1.058006
2000-01-02 -0.080372 -1.268121  1.561967
2000-01-03  0.816983  1.965656 -1.169408
2000-01-04  0.712795 -0.062433  0.736755
2000-01-05 -0.298721 -1.988045  1.475308
2000-01-06  1.103675  1.382242 -0.650762
2000-01-07 -0.729161 -0.142928 -1.063038
2000-01-08 -1.005977  0.465222 -0.094517

Deletion of the object specified by the key:

# store.remove('df') is an equivalent method
In [463]: del store["df"]

In [464]: store
Out[464]: 
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

Closing a Store and using a context manager:

In [465]: store.close()

In [466]: store
Out[466]: 
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

In [467]: store.is_open
Out[467]: False

# Working with, and automatically closing the store using a context manager
In [468]: with pd.HDFStore("store.h5") as store:
   .....:     store.keys()
   .....: 

Read/write API#

HDFStore supports a top-level API using read_hdf for reading and to_hdf for writing, similar to how read_csv and to_csv work.

In [469]: df_tl = pd.DataFrame({"A": list(range(5)), "B": list(range(5))})

In [470]: df_tl.to_hdf("store_tl.h5", "table", append=True)

In [471]: pd.read_hdf("store_tl.h5", "table", where=["index>2"])
Out[471]: 
   A  B
3  3  3
4  4  4

HDFStore will by default not drop rows that are all missing. This behavior can be changed by setting dropna=True.

In [472]: df_with_missing = pd.DataFrame(
   .....:     {
   .....:         "col1": [0, np.nan, 2],
   .....:         "col2": [1, np.nan, np.nan],
   .....:     }
   .....: )
   .....: 

In [473]: df_with_missing
Out[473]: 
   col1  col2
0   0.0   1.0
1   NaN   NaN
2   2.0   NaN

In [474]: df_with_missing.to_hdf("file.h5", "df_with_missing", format="table", mode="w")

In [475]: pd.read_hdf("file.h5", "df_with_missing")
Out[475]: 
   col1  col2
0   0.0   1.0
1   NaN   NaN
2   2.0   NaN

In [476]: df_with_missing.to_hdf(
   .....:     "file.h5", "df_with_missing", format="table", mode="w", dropna=True
   .....: )
   .....: 

In [477]: pd.read_hdf("file.h5", "df_with_missing")
Out[477]: 
   col1  col2
0   0.0   1.0
2   2.0   NaN

Fixed format#

The examples above show storing using put, which write the HDF5 to PyTables in a fixed array format, called the fixed format. These types of stores are not appendable once written (though you can simply remove them and rewrite). Nor are they queryable; they must be retrieved in their entirety. They also do not support dataframes with non-unique column names. The fixed format stores offer very fast writing and slightly faster reading than table stores. This format is specified by default when using put or to_hdf or by format='fixed' or format='f'.

Warning

A fixed format will raise a TypeError if you try to retrieve using a where:

In [478]: pd.DataFrame(np.random.randn(10, 2)).to_hdf("test_fixed.h5", "df")

In [479]: pd.read_hdf("test_fixed.h5", "df", where="index>5")
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[479], line 1
----> 1 pd.read_hdf("test_fixed.h5", "df", where="index>5")

File ~/work/pandas/pandas/pandas/io/pytables.py:451, in read_hdf(path_or_buf, key, mode, errors, where, start, stop, columns, iterator, chunksize, **kwargs)
    446                 raise ValueError(
    447                     "key must be provided when HDF5 "
    448                     "file contains multiple datasets."
    449                 )
    450         key = candidate_only_group._v_pathname
--> 451     return store.select(
    452         key,
    453         where=where,
    454         start=start,
    455         stop=stop,
    456         columns=columns,
    457         iterator=iterator,
    458         chunksize=chunksize,
    459         auto_close=auto_close,
    460     )
    461 except (ValueError, TypeError, LookupError):
    462     if not isinstance(path_or_buf, HDFStore):
    463         # if there is an error, close the store if we opened it.

File ~/work/pandas/pandas/pandas/io/pytables.py:905, in HDFStore.select(self, key, where, start, stop, columns, iterator, chunksize, auto_close)
    891 # create the iterator
    892 it = TableIterator(
    893     self,
    894     s,
   (...)
    902     auto_close=auto_close,
    903 )
--> 905 return it.get_result()

File ~/work/pandas/pandas/pandas/io/pytables.py:2028, in TableIterator.get_result(self, coordinates)
   2025     where = self.where
   2027 # directly return the result
-> 2028 results = self.func(self.start, self.stop, where)
   2029 self.close()
   2030 return results

File ~/work/pandas/pandas/pandas/io/pytables.py:889, in HDFStore.select.<locals>.func(_start, _stop, _where)
    888 def func(_start, _stop, _where):
--> 889     return s.read(start=_start, stop=_stop, where=_where, columns=columns)

File ~/work/pandas/pandas/pandas/io/pytables.py:3276, in BlockManagerFixed.read(self, where, columns, start, stop)
   3268 def read(
   3269     self,
   3270     where=None,
   (...)
   3274 ) -> DataFrame:
   3275     # start, stop applied to rows, so 0th axis only
-> 3276     self.validate_read(columns, where)
   3277     select_axis = self.obj_type()._get_block_manager_axis(0)
   3279     axes = []

File ~/work/pandas/pandas/pandas/io/pytables.py:2918, in GenericFixed.validate_read(self, columns, where)
   2913     raise TypeError(
   2914         "cannot pass a column specification when reading "
   2915         "a Fixed format store. this store must be selected in its entirety"
   2916     )
   2917 if where is not None:
-> 2918     raise TypeError(
   2919         "cannot pass a where specification when reading "
   2920         "from a Fixed format store. this store must be selected in its entirety"
   2921     )

TypeError: cannot pass a where specification when reading from a Fixed format store. this store must be selected in its entirety

Table format#

HDFStore supports another PyTables format on disk, the table format. Conceptually a table is shaped very much like a DataFrame, with rows and columns. A table may be appended to in the same or other sessions. In addition, delete and query type operations are supported. This format is specified by format='table' or format='t' to append or put or to_hdf.

This format can be set as an option as well pd.set_option('io.hdf.default_format','table') to enable put/append/to_hdf to by default store in the table format.

In [480]: store = pd.HDFStore("store.h5")

In [481]: df1 = df[0:4]

In [482]: df2 = df[4:]

# append data (creates a table automatically)
In [483]: store.append("df", df1)

In [484]: store.append("df", df2)

In [485]: store
Out[485]: 
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

# select the entire object
In [486]: store.select("df")
Out[486]: 
                   A         B         C
2000-01-01  0.858644 -0.851236  1.058006
2000-01-02 -0.080372 -1.268121  1.561967
2000-01-03  0.816983  1.965656 -1.169408
2000-01-04  0.712795 -0.062433  0.736755
2000-01-05 -0.298721 -1.988045  1.475308
2000-01-06  1.103675  1.382242 -0.650762
2000-01-07 -0.729161 -0.142928 -1.063038
2000-01-08 -1.005977  0.465222 -0.094517

# the type of stored data
In [487]: store.root.df._v_attrs.pandas_type
Out[487]: 'frame_table'

Note

You can also create a table by passing format='table' or format='t' to a put operation.

Hierarchical keys#

Keys to a store can be specified as a string. These can be in a hierarchical path-name like format (e.g. foo/bar/bah), which will generate a hierarchy of sub-stores (or Groups in PyTables parlance). Keys can be specified without the leading ‘/’ and are always absolute (e.g. ‘foo’ refers to ‘/foo’). Removal operations can remove everything in the sub-store and below, so be careful.

In [488]: store.put("foo/bar/bah", df)

In [489]: store.append("food/orange", df)

In [490]: store.append("food/apple", df)

In [491]: store
Out[491]: 
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

# a list of keys are returned
In [492]: store.keys()
Out[492]: ['/df', '/food/apple', '/food/orange', '/foo/bar/bah']

# remove all nodes under this level
In [493]: store.remove("food")

In [494]: store
Out[494]: 
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

You can walk through the group hierarchy using the walk method which will yield a tuple for each group key along with the relative keys of its contents.

In [495]: for (path, subgroups, subkeys) in store.walk():
   .....:     for subgroup in subgroups:
   .....:         print("GROUP: {}/{}".format(path, subgroup))
   .....:     for subkey in subkeys:
   .....:         key = "/".join([path, subkey])
   .....:         print("KEY: {}".format(key))
   .....:         print(store.get(key))
   .....: 
GROUP: /foo
KEY: /df
                   A         B         C
2000-01-01  0.858644 -0.851236  1.058006
2000-01-02 -0.080372 -1.268121  1.561967
2000-01-03  0.816983  1.965656 -1.169408
2000-01-04  0.712795 -0.062433  0.736755
2000-01-05 -0.298721 -1.988045  1.475308
2000-01-06  1.103675  1.382242 -0.650762
2000-01-07 -0.729161 -0.142928 -1.063038
2000-01-08 -1.005977  0.465222 -0.094517
GROUP: /foo/bar
KEY: /foo/bar/bah
                   A         B         C
2000-01-01  0.858644 -0.851236  1.058006
2000-01-02 -0.080372 -1.268121  1.561967
2000-01-03  0.816983  1.965656 -1.169408
2000-01-04  0.712795 -0.062433  0.736755
2000-01-05 -0.298721 -1.988045  1.475308
2000-01-06  1.103675  1.382242 -0.650762
2000-01-07 -0.729161 -0.142928 -1.063038
2000-01-08 -1.005977  0.465222 -0.094517

Warning

Hierarchical keys cannot be retrieved as dotted (attribute) access as described above for items stored under the root node.

In [496]: store.foo.bar.bah
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[496], line 1
----> 1 store.foo.bar.bah

File ~/work/pandas/pandas/pandas/io/pytables.py:612, in HDFStore.__getattr__(self, name)
    610 """allow attribute access to get stores"""
    611 try:
--> 612     return self.get(name)
    613 except (KeyError, ClosedFileError):
    614     pass

File ~/work/pandas/pandas/pandas/io/pytables.py:812, in HDFStore.get(self, key)
    810 if group is None:
    811     raise KeyError(f"No object named {key} in the file")
--> 812 return self._read_group(group)

File ~/work/pandas/pandas/pandas/io/pytables.py:1877, in HDFStore._read_group(self, group)
   1876 def _read_group(self, group: Node):
-> 1877     s = self._create_storer(group)
   1878     s.infer_axes()
   1879     return s.read()

File ~/work/pandas/pandas/pandas/io/pytables.py:1751, in HDFStore._create_storer(self, group, format, value, encoding, errors)
   1749         tt = "generic_table"
   1750     else:
-> 1751         raise TypeError(
   1752             "cannot create a storer if the object is not existing "
   1753             "nor a value are passed"
   1754         )
   1755 else:
   1756     if isinstance(value, Series):

TypeError: cannot create a storer if the object is not existing nor a value are passed
# you can directly access the actual PyTables node but using the root node
In [497]: store.root.foo.bar.bah
Out[497]: 
/foo/bar/bah (Group) ''
  children := ['axis0' (Array), 'axis1' (Array), 'block0_items' (Array), 'block0_values' (Array)]

Instead, use explicit string based keys:

In [498]: store["foo/bar/bah"]
Out[498]: 
                   A         B         C
2000-01-01  0.858644 -0.851236  1.058006
2000-01-02 -0.080372 -1.268121  1.561967
2000-01-03  0.816983  1.965656 -1.169408
2000-01-04  0.712795 -0.062433  0.736755
2000-01-05 -0.298721 -1.988045  1.475308
2000-01-06  1.103675  1.382242 -0.650762
2000-01-07 -0.729161 -0.142928 -1.063038
2000-01-08 -1.005977  0.465222 -0.094517

Storing types#

Storing mixed types in a table#

Storing mixed-dtype data is supported. Strings are stored as a fixed-width using the maximum size of the appended column. Subsequent attempts at appending longer strings will raise a ValueError.

Passing min_itemsize={`values`: size} as a parameter to append will set a larger minimum for the string columns. Storing floats, strings, ints, bools, datetime64 are currently supported. For string columns, passing nan_rep = 'nan' to append will change the default nan representation on disk (which converts to/from np.nan), this defaults to nan.

In [499]: df_mixed = pd.DataFrame(
   .....:     {
   .....:         "A": np.random.randn(8),
   .....:         "B": np.random.randn(8),
   .....:         "C": np.array(np.random.randn(8), dtype="float32"),
   .....:         "string": "string",
   .....:         "int": 1,
   .....:         "bool": True,
   .....:         "datetime64": pd.Timestamp("20010102"),
   .....:     },
   .....:     index=list(range(8)),
   .....: )
   .....: 

In [500]: df_mixed.loc[df_mixed.index[3:5], ["A", "B", "string", "datetime64"]] = np.nan

In [501]: store.append("df_mixed", df_mixed, min_itemsize={"values": 50})

In [502]: df_mixed1 = store.select("df_mixed")

In [503]: df_mixed1
Out[503]: 
          A         B         C  ... int  bool                    datetime64
0  0.013747 -1.166078 -1.292080  ...   1  True 1970-01-01 00:00:00.978393600
1 -0.712009  0.247572  1.526911  ...   1  True 1970-01-01 00:00:00.978393600
2 -0.645096  1.687406  0.288504  ...   1  True 1970-01-01 00:00:00.978393600
3       NaN       NaN  0.097771  ...   1  True                           NaT
4       NaN       NaN  1.536408  ...   1  True                           NaT
5 -0.023202  0.043702  0.926790  ...   1  True 1970-01-01 00:00:00.978393600
6  2.359782  0.088224 -0.676448  ...   1  True 1970-01-01 00:00:00.978393600
7 -0.143428 -0.813360 -0.179724  ...   1  True 1970-01-01 00:00:00.978393600

[8 rows x 7 columns]

In [504]: df_mixed1.dtypes.value_counts()
Out[504]: 
float64           2
float32           1
object            1
int64             1
bool              1
datetime64[ns]    1
Name: count, dtype: int64

# we have provided a minimum string column size
In [505]: store.root.df_mixed.table
Out[505]: 
/df_mixed/table (Table(8,)) ''
  description := {
  "index": Int64Col(shape=(), dflt=0, pos=0),
  "values_block_0": Float64Col(shape=(2,), dflt=0.0, pos=1),
  "values_block_1": Float32Col(shape=(1,), dflt=0.0, pos=2),
  "values_block_2": StringCol(itemsize=50, shape=(1,), dflt=b'', pos=3),
  "values_block_3": Int64Col(shape=(1,), dflt=0, pos=4),
  "values_block_4": BoolCol(shape=(1,), dflt=False, pos=5),
  "values_block_5": Int64Col(shape=(1,), dflt=0, pos=6)}
  byteorder := 'little'
  chunkshape := (689,)
  autoindex := True
  colindexes := {
    "index": Index(6, mediumshuffle, zlib(1)).is_csi=False}

Storing MultiIndex DataFrames#

Storing MultiIndex DataFrames as tables is very similar to storing/selecting from homogeneous index DataFrames.

In [506]: index = pd.MultiIndex(
   .....:    levels=[["foo", "bar", "baz", "qux"], ["one", "two", "three"]],
   .....:    codes=[[0, 0, 0, 1, 1, 2, 2, 3, 3, 3], [0, 1, 2, 0, 1, 1, 2, 0, 1, 2]],
   .....:    names=["foo", "bar"],
   .....: )
   .....: 

In [507]: df_mi = pd.DataFrame(np.random.randn(10, 3), index=index, columns=["A", "B", "C"])

In [508]: df_mi
Out[508]: 
                  A         B         C
foo bar                                
foo one   -1.303456 -0.642994 -0.649456
    two    1.012694  0.414147  1.950460
    three  1.094544 -0.802899 -0.583343
bar one    0.410395  0.618321  0.560398
    two    1.434027 -0.033270  0.343197
baz two   -1.646063 -0.695847 -0.429156
    three -0.244688 -1.428229 -0.138691
qux one    1.866184 -1.446617  0.036660
    two   -1.660522  0.929553 -1.298649
    three  3.565769  0.682402  1.041927

In [509]: store.append("df_mi", df_mi)

In [510]: store.select("df_mi")
Out[510]: 
                  A         B         C
foo bar                                
foo one   -1.303456 -0.642994 -0.649456
    two    1.012694  0.414147  1.950460
    three  1.094544 -0.802899 -0.583343
bar one    0.410395  0.618321  0.560398
    two    1.434027 -0.033270  0.343197
baz two   -1.646063 -0.695847 -0.429156
    three -0.244688 -1.428229 -0.138691
qux one    1.866184 -1.446617  0.036660
    two   -1.660522  0.929553 -1.298649
    three  3.565769  0.682402  1.041927

# the levels are automatically included as data columns
In [511]: store.select("df_mi", "foo=bar")
Out[511]: 
                A         B         C
foo bar                              
bar one  0.410395  0.618321  0.560398
    two  1.434027 -0.033270  0.343197

Note

The index keyword is reserved and cannot be use as a level name.

Querying#

Querying a table#

select and delete operations have an optional criterion that can be specified to select/delete only a subset of the data. This allows one to have a very large on-disk table and retrieve only a portion of the data.

A query is specified using the Term class under the hood, as a boolean expression.

  • index and columns are supported indexers of DataFrames.

  • if data_columns are specified, these can be used as additional indexers.

  • level name in a MultiIndex, with default name level_0, level_1, … if not provided.

Valid comparison operators are:

=, ==, !=, >, >=, <, <=

Valid boolean expressions are combined with:

  • | : or

  • & : and

  • ( and ) : for grouping

These rules are similar to how boolean expressions are used in pandas for indexing.

Note

  • = will be automatically expanded to the comparison operator ==

  • ~ is the not operator, but can only be used in very limited circumstances

  • If a list/tuple of expressions is passed they will be combined via &

The following are valid expressions:

  • 'index >= date'

  • "columns = ['A', 'D']"

  • "columns in ['A', 'D']"

  • 'columns = A'

  • 'columns == A'

  • "~(columns = ['A', 'B'])"

  • 'index > df.index[3] & string = "bar"'

  • '(index > df.index[3] & index <= df.index[6]) | string = "bar"'

  • "ts >= Timestamp('2012-02-01')"

  • "major_axis>=20130101"

The indexers are on the left-hand side of the sub-expression:

columns, major_axis, ts

The right-hand side of the sub-expression (after a comparison operator) can be:

  • functions that will be evaluated, e.g. Timestamp('2012-02-01')

  • strings, e.g. "bar"

  • date-like, e.g. 20130101, or "20130101"

  • lists, e.g. "['A', 'B']"

  • variables that are defined in the local names space, e.g. date

Note

Passing a string to a query by interpolating it into the query expression is not recommended. Simply assign the string of interest to a variable and use that variable in an expression. For example, do this

string = "HolyMoly'"
store.select("df", "index == string")

instead of this

string = "HolyMoly'"
store.select('df', f'index == {string}')

The latter will not work and will raise a SyntaxError.Note that there’s a single quote followed by a double quote in the string variable.

If you must interpolate, use the '%r' format specifier

store.select("df", "index == %r" % string)

which will quote string.

Here are some examples:

In [512]: dfq = pd.DataFrame(
   .....:     np.random.randn(10, 4),
   .....:     columns=list("ABCD"),
   .....:     index=pd.date_range("20130101", periods=10),
   .....: )
   .....: 

In [513]: store.append("dfq", dfq, format="table", data_columns=True)

Use boolean expressions, with in-line function evaluation.

In [514]: store.select("dfq", "index>pd.Timestamp('20130104') & columns=['A', 'B']")
Out[514]: 
                   A         B
2013-01-05 -0.830545 -0.457071
2013-01-06  0.431186  1.049421
2013-01-07  0.617509 -0.811230
2013-01-08  0.947422 -0.671233
2013-01-09 -0.183798 -1.211230
2013-01-10  0.361428  0.887304

Use inline column reference.

In [515]: store.select("dfq", where="A>0 or C>0")
Out[515]: 
                   A         B         C         D
2013-01-02  0.658179  0.362814 -0.917897  0.010165
2013-01-03  0.905122  1.848731 -1.184241  0.932053
2013-01-05 -0.830545 -0.457071  1.565581  1.148032
2013-01-06  0.431186  1.049421  0.383309  0.595013
2013-01-07  0.617509 -0.811230 -2.088563 -1.393500
2013-01-08  0.947422 -0.671233 -0.847097 -1.187785
2013-01-10  0.361428  0.887304  0.266457 -0.399641

The columns keyword can be supplied to select a list of columns to be returned, this is equivalent to passing a 'columns=list_of_columns_to_filter':

In [516]: store.select("df", "columns=['A', 'B']")
Out[516]: 
                   A         B
2000-01-01  0.858644 -0.851236
2000-01-02 -0.080372 -1.268121
2000-01-03  0.816983  1.965656
2000-01-04  0.712795 -0.062433
2000-01-05 -0.298721 -1.988045
2000-01-06  1.103675  1.382242
2000-01-07 -0.729161 -0.142928
2000-01-08 -1.005977  0.465222

start and stop parameters can be specified to limit the total search space. These are in terms of the total number of rows in a table.

Note

select will raise a ValueError if the query expression has an unknown variable reference. Usually this means that you are trying to select on a column that is not a data_column.

select will raise a SyntaxError if the query expression is not valid.

Query timedelta64[ns]#

You can store and query using the timedelta64[ns] type. Terms can be specified in the format: <float>(<unit>), where float may be signed (and fractional), and unit can be D,s,ms,us,ns for the timedelta. Here’s an example:

In [517]: from datetime import timedelta

In [518]: dftd = pd.DataFrame(
   .....:     {
   .....:         "A": pd.Timestamp("20130101"),
   .....:         "B": [
   .....:             pd.Timestamp("20130101") + timedelta(days=i, seconds=10)
   .....:             for i in range(10)
   .....:         ],
   .....:     }
   .....: )
   .....: 

In [519]: dftd["C"] = dftd["A"] - dftd["B"]

In [520]: dftd
Out[520]: 
           A                   B                  C
0 2013-01-01 2013-01-01 00:00:10  -1 days +23:59:50
1 2013-01-01 2013-01-02 00:00:10  -2 days +23:59:50
2 2013-01-01 2013-01-03 00:00:10  -3 days +23:59:50
3 2013-01-01 2013-01-04 00:00:10  -4 days +23:59:50
4 2013-01-01 2013-01-05 00:00:10  -5 days +23:59:50
5 2013-01-01 2013-01-06 00:00:10  -6 days +23:59:50
6 2013-01-01 2013-01-07 00:00:10  -7 days +23:59:50
7 2013-01-01 2013-01-08 00:00:10  -8 days +23:59:50
8 2013-01-01 2013-01-09 00:00:10  -9 days +23:59:50
9 2013-01-01 2013-01-10 00:00:10 -10 days +23:59:50

In [521]: store.append("dftd", dftd, data_columns=True)

In [522]: store.select("dftd", "C<'-3.5D'")
Out[522]: 
                              A                   B                  C
4 1970-01-01 00:00:01.356998400 2013-01-05 00:00:10  -5 days +23:59:50
5 1970-01-01 00:00:01.356998400 2013-01-06 00:00:10  -6 days +23:59:50
6 1970-01-01 00:00:01.356998400 2013-01-07 00:00:10  -7 days +23:59:50
7 1970-01-01 00:00:01.356998400 2013-01-08 00:00:10  -8 days +23:59:50
8 1970-01-01 00:00:01.356998400 2013-01-09 00:00:10  -9 days +23:59:50
9 1970-01-01 00:00:01.356998400 2013-01-10 00:00:10 -10 days +23:59:50

Query MultiIndex#

Selecting from a MultiIndex can be achieved by using the name of the level.

In [523]: df_mi.index.names
Out[523]: FrozenList(['foo', 'bar'])

In [524]: store.select("df_mi", "foo=baz and bar=two")
Out[524]: 
                A         B         C
foo bar                              
baz two -1.646063 -0.695847 -0.429156

If the MultiIndex levels names are None, the levels are automatically made available via the level_n keyword with n the level of the MultiIndex you want to select from.

In [525]: index = pd.MultiIndex(
   .....:     levels=[["foo", "bar", "baz", "qux"], ["one", "two", "three"]],
   .....:     codes=[[0, 0, 0, 1, 1, 2, 2, 3, 3, 3], [0, 1, 2, 0, 1, 1, 2, 0, 1, 2]],
   .....: )
   .....: 

In [526]: df_mi_2 = pd.DataFrame(np.random.randn(10, 3), index=index, columns=["A", "B", "C"])

In [527]: df_mi_2
Out[527]: 
                  A         B         C
foo one   -0.219582  1.186860 -1.437189
    two    0.053768  1.872644 -1.469813
    three -0.564201  0.876341  0.407749
bar one   -0.232583  0.179812  0.922152
    two   -1.820952 -0.641360  2.133239
baz two   -0.941248 -0.136307 -1.271305
    three -0.099774 -0.061438 -0.845172
qux one    0.465793  0.756995 -0.541690
    two   -0.802241  0.877657 -2.553831
    three  0.094899 -2.319519  0.293601

In [528]: store.append("df_mi_2", df_mi_2)

# the levels are automatically included as data columns with keyword level_n
In [529]: store.select("df_mi_2", "level_0=foo and level_1=two")
Out[529]: 
                A         B         C
foo two  0.053768  1.872644 -1.469813

Indexing#

You can create/modify an index for a table with create_table_index after data is already in the table (after and append/put operation). Creating a table index is highly encouraged. This will speed your queries a great deal when you use a select with the indexed dimension as the where.

Note

Indexes are automagically created on the indexables and any data columns you specify. This behavior can be turned off by passing index=False to append.

# we have automagically already created an index (in the first section)
In [530]: i = store.root.df.table.cols.index.index

In [531]: i.optlevel, i.kind
Out[531]: (6, 'medium')

# change an index by passing new parameters
In [532]: store.create_table_index("df", optlevel=9, kind="full")

In [533]: i = store.root.df.table.cols.index.index

In [534]: i.optlevel, i.kind
Out[534]: (9, 'full')

Oftentimes when appending large amounts of data to a store, it is useful to turn off index creation for each append, then recreate at the end.

In [535]: df_1 = pd.DataFrame(np.random.randn(10, 2), columns=list("AB"))

In [536]: df_2 = pd.DataFrame(np.random.randn(10, 2), columns=list("AB"))

In [537]: st = pd.HDFStore("appends.h5", mode="w")

In [538]: st.append("df", df_1, data_columns=["B"], index=False)

In [539]: st.append("df", df_2, data_columns=["B"], index=False)

In [540]: st.get_storer("df").table
Out[540]: 
/df/table (Table(20,)) ''
  description := {
  "index": Int64Col(shape=(), dflt=0, pos=0),
  "values_block_0": Float64Col(shape=(1,), dflt=0.0, pos=1),
  "B": Float64Col(shape=(), dflt=0.0, pos=2)}
  byteorder := 'little'
  chunkshape := (2730,)

Then create the index when finished appending.

In [541]: st.create_table_index("df", columns=["B"], optlevel=9, kind="full")

In [542]: st.get_storer("df").table
Out[542]: 
/df/table (Table(20,)) ''
  description := {
  "index": Int64Col(shape=(), dflt=0, pos=0),
  "values_block_0": Float64Col(shape=(1,), dflt=0.0, pos=1),
  "B": Float64Col(shape=(), dflt=0.0, pos=2)}
  byteorder := 'little'
  chunkshape := (2730,)
  autoindex := True
  colindexes := {
    "B": Index(9, fullshuffle, zlib(1)).is_csi=True}

In [543]: st.close()

See here for how to create a completely-sorted-index (CSI) on an existing store.

Query via data columns#

You can designate (and index) certain columns that you want to be able to perform queries (other than the indexable columns, which you can always query). For instance say you want to perform this common operation, on-disk, and return just the frame that matches this query. You can specify data_columns = True to force all columns to be data_columns.

In [544]: df_dc = df.copy()

In [545]: df_dc["string"] = "foo"

In [546]: df_dc.loc[df_dc.index[4:6], "string"] = np.nan

In [547]: df_dc.loc[df_dc.index[7:9], "string"] = "bar"

In [548]: df_dc["string2"] = "cool"

In [549]: df_dc.loc[df_dc.index[1:3], ["B", "C"]] = 1.0

In [550]: df_dc
Out[550]: 
                   A         B         C string string2
2000-01-01  0.858644 -0.851236  1.058006    foo    cool
2000-01-02 -0.080372  1.000000  1.000000    foo    cool
2000-01-03  0.816983  1.000000  1.000000    foo    cool
2000-01-04  0.712795 -0.062433  0.736755    foo    cool
2000-01-05 -0.298721 -1.988045  1.475308    NaN    cool
2000-01-06  1.103675  1.382242 -0.650762    NaN    cool
2000-01-07 -0.729161 -0.142928 -1.063038    foo    cool
2000-01-08 -1.005977  0.465222 -0.094517    bar    cool

# on-disk operations
In [551]: store.append("df_dc", df_dc, data_columns=["B", "C", "string", "string2"])

In [552]: store.select("df_dc", where="B > 0")
Out[552]: 
                   A         B         C string string2
2000-01-02 -0.080372  1.000000  1.000000    foo    cool
2000-01-03  0.816983  1.000000  1.000000    foo    cool
2000-01-06  1.103675  1.382242 -0.650762    NaN    cool
2000-01-08 -1.005977  0.465222 -0.094517    bar    cool

# getting creative
In [553]: store.select("df_dc", "B > 0 & C > 0 & string == foo")
Out[553]: 
                   A    B    C string string2
2000-01-02 -0.080372  1.0  1.0    foo    cool
2000-01-03  0.816983  1.0  1.0    foo    cool

# this is in-memory version of this type of selection
In [554]: df_dc[(df_dc.B > 0) & (df_dc.C > 0) & (df_dc.string == "foo")]
Out[554]: 
                   A    B    C string string2
2000-01-02 -0.080372  1.0  1.0    foo    cool
2000-01-03  0.816983  1.0  1.0    foo    cool

# we have automagically created this index and the B/C/string/string2
# columns are stored separately as ``PyTables`` columns
In [555]: store.root.df_dc.table
Out[555]: 
/df_dc/table (Table(8,)) ''
  description := {
  "index": Int64Col(shape=(), dflt=0, pos=0),
  "values_block_0": Float64Col(shape=(1,), dflt=0.0, pos=1),
  "B": Float64Col(shape=(), dflt=0.0, pos=2),
  "C": Float64Col(shape=(), dflt=0.0, pos=3),
  "string": StringCol(itemsize=3, shape=(), dflt=b'', pos=4),
  "string2": StringCol(itemsize=4, shape=(), dflt=b'', pos=5)}
  byteorder := 'little'
  chunkshape := (1680,)
  autoindex := True
  colindexes := {
    "index": Index(6, mediumshuffle, zlib(1)).is_csi=False,
    "B": Index(6, mediumshuffle, zlib(1)).is_csi=False,
    "C": Index(6, mediumshuffle, zlib(1)).is_csi=False,
    "string": Index(6, mediumshuffle, zlib(1)).is_csi=False,
    "string2": Index(6, mediumshuffle, zlib(1)).is_csi=False}

There is some performance degradation by making lots of columns into data columns, so it is up to the user to designate these. In addition, you cannot change data columns (nor indexables) after the first append/put operation (Of course you can simply read in the data and create a new table!).

Iterator#

You can pass iterator=True or chunksize=number_in_a_chunk to select and select_as_multiple to return an iterator on the results. The default is 50,000 rows returned in a chunk.

In [556]: for df in store.select("df", chunksize=3):
   .....:     print(df)
   .....: 
                   A         B         C
2000-01-01  0.858644 -0.851236  1.058006
2000-01-02 -0.080372 -1.268121  1.561967
2000-01-03  0.816983  1.965656 -1.169408
                   A         B         C
2000-01-04  0.712795 -0.062433  0.736755
2000-01-05 -0.298721 -1.988045  1.475308
2000-01-06  1.103675  1.382242 -0.650762
                   A         B         C
2000-01-07 -0.729161 -0.142928 -1.063038
2000-01-08 -1.005977  0.465222 -0.094517

Note

You can also use the iterator with read_hdf which will open, then automatically close the store when finished iterating.

for df in pd.read_hdf("store.h5", "df", chunksize=3):
    print(df)

Note, that the chunksize keyword applies to the source rows. So if you are doing a query, then the chunksize will subdivide the total rows in the table and the query applied, returning an iterator on potentially unequal sized chunks.

Here is a recipe for generating a query and using it to create equal sized return chunks.

In [557]: dfeq = pd.DataFrame({"number": np.arange(1, 11)})

In [558]: dfeq
Out[558]: 
   number
0       1
1       2
2       3
3       4
4       5
5       6
6       7
7       8
8       9
9      10

In [559]: store.append("dfeq", dfeq, data_columns=["number"])

In [560]: def chunks(l, n):
   .....:     return [l[i: i + n] for i in range(0, len(l), n)]
   .....: 

In [561]: evens = [2, 4, 6, 8, 10]

In [562]: coordinates = store.select_as_coordinates("dfeq", "number=evens")

In [563]: for c in chunks(coordinates, 2):
   .....:     print(store.select("dfeq", where=c))
   .....: 
   number
1       2
3       4
   number
5       6
7       8
   number
9      10

Advanced queries#

Select a single column#

To retrieve a single indexable or data column, use the method select_column. This will, for example, enable you to get the index very quickly. These return a Series of the result, indexed by the row number. These do not currently accept the where selector.

In [564]: store.select_column("df_dc", "index")
Out[564]: 
0   2000-01-01
1   2000-01-02
2   2000-01-03
3   2000-01-04
4   2000-01-05
5   2000-01-06
6   2000-01-07
7   2000-01-08
Name: index, dtype: datetime64[ns]

In [565]: store.select_column("df_dc", "string")
Out[565]: 
0    foo
1    foo
2    foo
3    foo
4    NaN
5    NaN
6    foo
7    bar
Name: string, dtype: object
Selecting coordinates#

Sometimes you want to get the coordinates (a.k.a the index locations) of your query. This returns an Index of the resulting locations. These coordinates can also be passed to subsequent where operations.

In [566]: df_coord = pd.DataFrame(
   .....:     np.random.randn(1000, 2), index=pd.date_range("20000101", periods=1000)
   .....: )
   .....: 

In [567]: store.append("df_coord", df_coord)

In [568]: c = store.select_as_coordinates("df_coord", "index > 20020101")

In [569]: c
Out[569]: 
Index([732, 733, 734, 735, 736, 737, 738, 739, 740, 741,
       ...
       990, 991, 992, 993, 994, 995, 996, 997, 998, 999],
      dtype='int64', length=268)

In [570]: store.select("df_coord", where=c)
Out[570]: 
                   0         1
2002-01-02  0.007717  1.168386
2002-01-03  0.759328 -0.638934
2002-01-04 -1.154018 -0.324071
2002-01-05 -0.804551 -1.280593
2002-01-06 -0.047208  1.260503
...              ...       ...
2002-09-22 -1.139583  0.344316
2002-09-23 -0.760643 -1.306704
2002-09-24  0.059018  1.775482
2002-09-25  1.242255 -0.055457
2002-09-26  0.410317  2.194489

[268 rows x 2 columns]
Selecting using a where mask#

Sometime your query can involve creating a list of rows to select. Usually this mask would be a resulting index from an indexing operation. This example selects the months of a datetimeindex which are 5.

In [571]: df_mask = pd.DataFrame(
   .....:     np.random.randn(1000, 2), index=pd.date_range("20000101", periods=1000)
   .....: )
   .....: 

In [572]: store.append("df_mask", df_mask)

In [573]: c = store.select_column("df_mask", "index")

In [574]: where = c[pd.DatetimeIndex(c).month == 5].index

In [575]: store.select("df_mask", where=where)
Out[575]: 
                   0         1
2000-05-01  1.479511  0.516433
2000-05-02 -0.334984 -1.493537
2000-05-03  0.900321  0.049695
2000-05-04  0.614266 -1.077151
2000-05-05  0.233881  0.493246
...              ...       ...
2002-05-27  0.294122  0.457407
2002-05-28 -1.102535  1.215650
2002-05-29 -0.432911  0.753606
2002-05-30 -1.105212  2.311877
2002-05-31  2.567296  2.610691

[93 rows x 2 columns]
Storer object#

If you want to inspect the stored object, retrieve via get_storer. You could use this programmatically to say get the number of rows in an object.

In [576]: store.get_storer("df_dc").nrows
Out[576]: 8

Multiple table queries#

The methods append_to_multiple and select_as_multiple can perform appending/selecting from multiple tables at once. The idea is to have one table (call it the selector table) that you index most/all of the columns, and perform your queries. The other table(s) are data tables with an index matching the selector table’s index. You can then perform a very fast query on the selector table, yet get lots of data back. This method is similar to having a very wide table, but enables more efficient queries.

The append_to_multiple method splits a given single DataFrame into multiple tables according to d, a dictionary that maps the table names to a list of ‘columns’ you want in that table. If None is used in place of a list, that table will have the remaining unspecified columns of the given DataFrame. The argument selector defines which table is the selector table (which you can make queries from). The argument dropna will drop rows from the input DataFrame to ensure tables are synchronized. This means that if a row for one of the tables being written to is entirely np.nan, that row will be dropped from all tables.

If dropna is False, THE USER IS RESPONSIBLE FOR SYNCHRONIZING THE TABLES. Remember that entirely np.Nan rows are not written to the HDFStore, so if you choose to call dropna=False, some tables may have more rows than others, and therefore select_as_multiple may not work or it may return unexpected results.

In [577]: df_mt = pd.DataFrame(
   .....:     np.random.randn(8, 6),
   .....:     index=pd.date_range("1/1/2000", periods=8),
   .....:     columns=["A", "B", "C", "D", "E", "F"],
   .....: )
   .....: 

In [578]: df_mt["foo"] = "bar"

In [579]: df_mt.loc[df_mt.index[1], ("A", "B")] = np.nan

# you can also create the tables individually
In [580]: store.append_to_multiple(
   .....:     {"df1_mt": ["A", "B"], "df2_mt": None}, df_mt, selector="df1_mt"
   .....: )
   .....: 

In [581]: store
Out[581]: 
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5

# individual tables were created
In [582]: store.select("df1_mt")
Out[582]: 
                   A         B
2000-01-01  0.162291 -0.430489
2000-01-02       NaN       NaN
2000-01-03  0.429207 -1.099274
2000-01-04  1.869081 -1.466039
2000-01-05  0.092130 -1.726280
2000-01-06  0.266901 -0.036854
2000-01-07 -0.517871 -0.990317
2000-01-08 -0.231342  0.557402

In [583]: store.select("df2_mt")
Out[583]: 
                   C         D         E         F  foo
2000-01-01 -2.502042  0.668149  0.460708  1.834518  bar
2000-01-02  0.130441 -0.608465  0.439872  0.506364  bar
2000-01-03 -1.069546  1.236277  0.116634 -1.772519  bar
2000-01-04  0.137462  0.313939  0.748471 -0.943009  bar
2000-01-05  0.836517  2.049798  0.562167  0.189952  bar
2000-01-06  1.112750 -0.151596  1.503311  0.939470  bar
2000-01-07 -0.294348  0.335844 -0.794159  1.495614  bar
2000-01-08  0.860312 -0.538674 -0.541986 -1.759606  bar

# as a multiple
In [584]: store.select_as_multiple(
   .....:     ["df1_mt", "df2_mt"],
   .....:     where=["A>0", "B>0"],
   .....:     selector="df1_mt",
   .....: )
   .....: 
Out[584]: 
Empty DataFrame
Columns: [A, B, C, D, E, F, foo]
Index: []

Delete from a table#

You can delete from a table selectively by specifying a where. In deleting rows, it is important to understand the PyTables deletes rows by erasing the rows, then moving the following data. Thus deleting can potentially be a very expensive operation depending on the orientation of your data. To get optimal performance, it’s worthwhile to have the dimension you are deleting be the first of the indexables.

Data is ordered (on the disk) in terms of the indexables. Here’s a simple use case. You store panel-type data, with dates in the major_axis and ids in the minor_axis. The data is then interleaved like this:

  • date_1
    • id_1

    • id_2

    • .

    • id_n

  • date_2
    • id_1

    • .

    • id_n

It should be clear that a delete operation on the major_axis will be fairly quick, as one chunk is removed, then the following data moved. On the other hand a delete operation on the minor_axis will be very expensive. In this case it would almost certainly be faster to rewrite the table using a where that selects all but the missing data.

Warning

Please note that HDF5 DOES NOT RECLAIM SPACE in the h5 files automatically. Thus, repeatedly deleting (or removing nodes) and adding again, WILL TEND TO INCREASE THE FILE SIZE.

To repack and clean the file, use ptrepack.

Notes & caveats#

Compression#

PyTables allows the stored data to be compressed. This applies to all kinds of stores, not just tables. Two parameters are used to control compression: complevel and complib.

  • complevel specifies if and how hard data is to be compressed. complevel=0 and complevel=None disables compression and 0<complevel<10 enables compression.

  • complib specifies which compression library to use. If nothing is specified the default library zlib is used. A compression library usually optimizes for either good compression rates or speed and the results will depend on the type of data. Which type of compression to choose depends on your specific needs and data. The list of supported compression libraries:

    • zlib: The default compression library. A classic in terms of compression, achieves good compression rates but is somewhat slow.

    • lzo: Fast compression and decompression.

    • bzip2: Good compression rates.

    • blosc: Fast compression and decompression.

      Support for alternative blosc compressors:

      • blosc:blosclz This is the default compressor for blosc

      • blosc:lz4: A compact, very popular and fast compressor.

      • blosc:lz4hc: A tweaked version of LZ4, produces better compression ratios at the expense of speed.

      • blosc:snappy: A popular compressor used in many places.

      • blosc:zlib: A classic; somewhat slower than the previous ones, but achieving better compression ratios.

      • blosc:zstd: An extremely well balanced codec; it provides the best compression ratios among the others above, and at reasonably fast speed.

    If complib is defined as something other than the listed libraries a ValueError exception is issued.

Note

If the library specified with the complib option is missing on your platform, compression defaults to zlib without further ado.

Enable compression for all objects within the file:

store_compressed = pd.HDFStore(
    "store_compressed.h5", complevel=9, complib="blosc:blosclz"
)

Or on-the-fly compression (this only applies to tables) in stores where compression is not enabled:

store.append("df", df, complib="zlib", complevel=5)

ptrepack#

PyTables offers better write performance when tables are compressed after they are written, as opposed to turning on compression at the very beginning. You can use the supplied PyTables utility ptrepack. In addition, ptrepack can change compression levels after the fact.

ptrepack --chunkshape=auto --propindexes --complevel=9 --complib=blosc in.h5 out.h5

Furthermore ptrepack in.h5 out.h5 will repack the file to allow you to reuse previously deleted space. Alternatively, one can simply remove the file and write again, or use the copy method.

Caveats#

Warning

HDFStore is not-threadsafe for writing. The underlying PyTables only supports concurrent reads (via threading or processes). If you need reading and writing at the same time, you need to serialize these operations in a single thread in a single process. You will corrupt your data otherwise. See the (GH 2397) for more information.

  • If you use locks to manage write access between multiple processes, you may want to use fsync() before releasing write locks. For convenience you can use store.flush(fsync=True) to do this for you.

  • Once a table is created columns (DataFrame) are fixed; only exactly the same columns can be appended

  • Be aware that timezones (e.g., pytz.timezone('US/Eastern')) are not necessarily equal across timezone versions. So if data is localized to a specific timezone in the HDFStore using one version of a timezone library and that data is updated with another version, the data will be converted to UTC since these timezones are not considered equal. Either use the same version of timezone library or use tz_convert with the updated timezone definition.

Warning

PyTables will show a NaturalNameWarning if a column name cannot be used as an attribute selector. Natural identifiers contain only letters, numbers, and underscores, and may not begin with a number. Other identifiers cannot be used in a where clause and are generally a bad idea.

DataTypes#

HDFStore will map an object dtype to the PyTables underlying dtype. This means the following types are known to work:

Type

Represents missing values

floating : float64, float32, float16

np.nan

integer : int64, int32, int8, uint64,uint32, uint8

boolean

datetime64[ns]

NaT

timedelta64[ns]

NaT

categorical : see the section below

object : strings

np.nan

unicode columns are not supported, and WILL FAIL.

Categorical data#

You can write data that contains category dtypes to a HDFStore. Queries work the same as if it was an object array. However, the category dtyped data is stored in a more efficient manner.

In [585]: dfcat = pd.DataFrame(
   .....:     {"A": pd.Series(list("aabbcdba")).astype("category"), "B": np.random.randn(8)}
   .....: )
   .....: 

In [586]: dfcat
Out[586]: 
   A         B
0  a -1.520478
1  a -1.069391
2  b -0.551981
3  b  0.452407
4  c  0.409257
5  d  0.301911
6  b -0.640843
7  a -2.253022

In [587]: dfcat.dtypes
Out[587]: 
A    category
B     float64
dtype: object

In [588]: cstore = pd.HDFStore("cats.h5", mode="w")

In [589]: cstore.append("dfcat", dfcat, format="table", data_columns=["A"])

In [590]: result = cstore.select("dfcat", where="A in ['b', 'c']")

In [591]: result
Out[591]: 
   A         B
2  b -0.551981
3  b  0.452407
4  c  0.409257
6  b -0.640843

In [592]: result.dtypes
Out[592]: 
A    category
B     float64
dtype: object

String columns#

min_itemsize

The underlying implementation of HDFStore uses a fixed column width (itemsize) for string columns. A string column itemsize is calculated as the maximum of the length of data (for that column) that is passed to the HDFStore, in the first append. Subsequent appends, may introduce a string for a column larger than the column can hold, an Exception will be raised (otherwise you could have a silent truncation of these columns, leading to loss of information). In the future we may relax this and allow a user-specified truncation to occur.

Pass min_itemsize on the first table creation to a-priori specify the minimum length of a particular string column. min_itemsize can be an integer, or a dict mapping a column name to an integer. You can pass values as a key to allow all indexables or data_columns to have this min_itemsize.

Passing a min_itemsize dict will cause all passed columns to be created as data_columns automatically.

Note

If you are not passing any data_columns, then the min_itemsize will be the maximum of the length of any string passed

In [593]: dfs = pd.DataFrame({"A": "foo", "B": "bar"}, index=list(range(5)))

In [594]: dfs
Out[594]: 
     A    B
0  foo  bar
1  foo  bar
2  foo  bar
3  foo  bar
4  foo  bar

# A and B have a size of 30
In [595]: store.append("dfs", dfs, min_itemsize=30)

In [596]: store.get_storer("dfs").table
Out[596]: 
/dfs/table (Table(5,)) ''
  description := {
  "index": Int64Col(shape=(), dflt=0, pos=0),
  "values_block_0": StringCol(itemsize=30, shape=(2,), dflt=b'', pos=1)}
  byteorder := 'little'
  chunkshape := (963,)
  autoindex := True
  colindexes := {
    "index": Index(6, mediumshuffle, zlib(1)).is_csi=False}

# A is created as a data_column with a size of 30
# B is size is calculated
In [597]: store.append("dfs2", dfs, min_itemsize={"A": 30})

In [598]: store.get_storer("dfs2").table
Out[598]: 
/dfs2/table (Table(5,)) ''
  description := {
  "index": Int64Col(shape=(), dflt=0, pos=0),
  "values_block_0": StringCol(itemsize=3, shape=(1,), dflt=b'', pos=1),
  "A": StringCol(itemsize=30, shape=(), dflt=b'', pos=2)}
  byteorder := 'little'
  chunkshape := (1598,)
  autoindex := True
  colindexes := {
    "index": Index(6, mediumshuffle, zlib(1)).is_csi=False,
    "A": Index(6, mediumshuffle, zlib(1)).is_csi=False}

nan_rep

String columns will serialize a np.nan (a missing value) with the nan_rep string representation. This defaults to the string value nan. You could inadvertently turn an actual nan value into a missing value.

In [599]: dfss = pd.DataFrame({"A": ["foo", "bar", "nan"]})

In [600]: dfss
Out[600]: 
     A
0  foo
1  bar
2  nan

In [601]: store.append("dfss", dfss)

In [602]: store.select("dfss")
Out[602]: 
     A
0  foo
1  bar
2  NaN

# here you need to specify a different nan rep
In [603]: store.append("dfss2", dfss, nan_rep="_nan_")

In [604]: store.select("dfss2")
Out[604]: 
     A
0  foo
1  bar
2  nan

Performance#

  • tables format come with a writing performance penalty as compared to fixed stores. The benefit is the ability to append/delete and query (potentially very large amounts of data). Write times are generally longer as compared with regular stores. Query times can be quite fast, especially on an indexed axis.

  • You can pass chunksize=<int> to append, specifying the write chunksize (default is 50000). This will significantly lower your memory usage on writing.

  • You can pass expectedrows=<int> to the first append, to set the TOTAL number of rows that PyTables will expect. This will optimize read/write performance.

  • Duplicate rows can be written to tables, but are filtered out in selection (with the last items being selected; thus a table is unique on major, minor pairs)

  • A PerformanceWarning will be raised if you are attempting to store types that will be pickled by PyTables (rather than stored as endemic types). See Here for more information and some solutions.

Feather#

Feather provides binary columnar serialization for data frames. It is designed to make reading and writing data frames efficient, and to make sharing data across data analysis languages easy.

Feather is designed to faithfully serialize and de-serialize DataFrames, supporting all of the pandas dtypes, including extension dtypes such as categorical and datetime with tz.

Several caveats:

  • The format will NOT write an Index, or MultiIndex for the DataFrame and will raise an error if a non-default one is provided. You can .reset_index() to store the index or .reset_index(drop=True) to ignore it.

  • Duplicate column names and non-string columns names are not supported

  • Actual Python objects in object dtype columns are not supported. These will raise a helpful error message on an attempt at serialization.

See the Full Documentation.

In [605]: df = pd.DataFrame(
   .....:     {
   .....:         "a": list("abc"),
   .....:         "b": list(range(1, 4)),
   .....:         "c": np.arange(3, 6).astype("u1"),
   .....:         "d": np.arange(4.0, 7.0, dtype="float64"),
   .....:         "e": [True, False, True],
   .....:         "f": pd.Categorical(list("abc")),
   .....:         "g": pd.date_range("20130101", periods=3),
   .....:         "h": pd.date_range("20130101", periods=3, tz="US/Eastern"),
   .....:         "i": pd.date_range("20130101", periods=3, freq="ns"),
   .....:     }
   .....: )
   .....: 

In [606]: df
Out[606]: 
   a  b  c  ...          g                         h                             i
0  a  1  3  ... 2013-01-01 2013-01-01 00:00:00-05:00 2013-01-01 00:00:00.000000000
1  b  2  4  ... 2013-01-02 2013-01-02 00:00:00-05:00 2013-01-01 00:00:00.000000001
2  c  3  5  ... 2013-01-03 2013-01-03 00:00:00-05:00 2013-01-01 00:00:00.000000002

[3 rows x 9 columns]

In [607]: df.dtypes
Out[607]: 
a                        object
b                         int64
c                         uint8
d                       float64
e                          bool
f                      category
g                datetime64[ns]
h    datetime64[ns, US/Eastern]
i                datetime64[ns]
dtype: object

Write to a feather file.

In [608]: df.to_feather("example.feather")

Read from a feather file.

In [609]: result = pd.read_feather("example.feather")

In [610]: result
Out[610]: 
   a  b  c  ...          g                         h                             i
0  a  1  3  ... 2013-01-01 2013-01-01 00:00:00-05:00 2013-01-01 00:00:00.000000000
1  b  2  4  ... 2013-01-02 2013-01-02 00:00:00-05:00 2013-01-01 00:00:00.000000001
2  c  3  5  ... 2013-01-03 2013-01-03 00:00:00-05:00 2013-01-01 00:00:00.000000002

[3 rows x 9 columns]

# we preserve dtypes
In [611]: result.dtypes
Out[611]: 
a                        object
b                         int64
c                         uint8
d                       float64
e                          bool
f                      category
g                datetime64[ns]
h    datetime64[ns, US/Eastern]
i                datetime64[ns]
dtype: object

Parquet#

Apache Parquet provides a partitioned binary columnar serialization for data frames. It is designed to make reading and writing data frames efficient, and to make sharing data across data analysis languages easy. Parquet can use a variety of compression techniques to shrink the file size as much as possible while still maintaining good read performance.

Parquet is designed to faithfully serialize and de-serialize DataFrame s, supporting all of the pandas dtypes, including extension dtypes such as datetime with tz.

Several caveats.

  • Duplicate column names and non-string columns names are not supported.

  • The pyarrow engine always writes the index to the output, but fastparquet only writes non-default indexes. This extra column can cause problems for non-pandas consumers that are not expecting it. You can force including or omitting indexes with the index argument, regardless of the underlying engine.

  • Index level names, if specified, must be strings.

  • In the pyarrow engine, categorical dtypes for non-string types can be serialized to parquet, but will de-serialize as their primitive dtype.

  • The pyarrow engine preserves the ordered flag of categorical dtypes with string types. fastparquet does not preserve the ordered flag.

  • Non supported types include Interval and actual Python object types. These will raise a helpful error message on an attempt at serialization. Period type is supported with pyarrow >= 0.16.0.

  • The pyarrow engine preserves extension data types such as the nullable integer and string data type (requiring pyarrow >= 0.16.0, and requiring the extension type to implement the needed protocols, see the extension types documentation).

You can specify an engine to direct the serialization. This can be one of pyarrow, or fastparquet, or auto. If the engine is NOT specified, then the pd.options.io.parquet.engine option is checked; if this is also auto, then pyarrow is tried, and falling back to fastparquet.

See the documentation for pyarrow and fastparquet.

Note

These engines are very similar and should read/write nearly identical parquet format files. pyarrow>=8.0.0 supports timedelta data, fastparquet>=0.1.4 supports timezone aware datetimes. These libraries differ by having different underlying dependencies (fastparquet by using numba, while pyarrow uses a c-library).

In [612]: df = pd.DataFrame(
   .....:     {
   .....:         "a": list("abc"),
   .....:         "b": list(range(1, 4)),
   .....:         "c": np.arange(3, 6).astype("u1"),
   .....:         "d": np.arange(4.0, 7.0, dtype="float64"),
   .....:         "e": [True, False, True],
   .....:         "f": pd.date_range("20130101", periods=3),
   .....:         "g": pd.date_range("20130101", periods=3, tz="US/Eastern"),
   .....:         "h": pd.Categorical(list("abc")),
   .....:         "i": pd.Categorical(list("abc"), ordered=True),
   .....:     }
   .....: )
   .....: 

In [613]: df
Out[613]: 
   a  b  c    d      e          f                         g  h  i
0  a  1  3  4.0   True 2013-01-01 2013-01-01 00:00:00-05:00  a  a
1  b  2  4  5.0  False 2013-01-02 2013-01-02 00:00:00-05:00  b  b
2  c  3  5  6.0   True 2013-01-03 2013-01-03 00:00:00-05:00  c  c

In [614]: df.dtypes
Out[614]: 
a                        object
b                         int64
c                         uint8
d                       float64
e                          bool
f                datetime64[ns]
g    datetime64[ns, US/Eastern]
h                      category
i                      category
dtype: object

Write to a parquet file.

In [615]: df.to_parquet("example_pa.parquet", engine="pyarrow")

In [616]: df.to_parquet("example_fp.parquet", engine="fastparquet")

Read from a parquet file.

In [617]: result = pd.read_parquet("example_fp.parquet", engine="fastparquet")

In [618]: result = pd.read_parquet("example_pa.parquet", engine="pyarrow")

In [619]: result.dtypes
Out[619]: 
a                        object
b                         int64
c                         uint8
d                       float64
e                          bool
f                datetime64[ns]
g    datetime64[ns, US/Eastern]
h                      category
i                      category
dtype: object

By setting the dtype_backend argument you can control the default dtypes used for the resulting DataFrame.

In [620]: result = pd.read_parquet("example_pa.parquet", engine="pyarrow", dtype_backend="pyarrow")

In [621]: result.dtypes
Out[621]: 
a                                      string[pyarrow]
b                                       int64[pyarrow]
c                                       uint8[pyarrow]
d                                      double[pyarrow]
e                                        bool[pyarrow]
f                               timestamp[ns][pyarrow]
g                timestamp[ns, tz=US/Eastern][pyarrow]
h    dictionary<values=string, indices=int32, order...
i    dictionary<values=string, indices=int32, order...
dtype: object

Note

Note that this is not supported for fastparquet.

Read only certain columns of a parquet file.

In [622]: result = pd.read_parquet(
   .....:     "example_fp.parquet",
   .....:     engine="fastparquet",
   .....:     columns=["a", "b"],
   .....: )
   .....: 

In [623]: result = pd.read_parquet(
   .....:     "example_pa.parquet",
   .....:     engine="pyarrow",
   .....:     columns=["a", "b"],
   .....: )
   .....: 

In [624]: result.dtypes
Out[624]: 
a    object
b     int64
dtype: object

Handling indexes#

Serializing a DataFrame to parquet may include the implicit index as one or more columns in the output file. Thus, this code:

In [625]: df = pd.DataFrame({"a": [1, 2], "b": [3, 4]})

In [626]: df.to_parquet("test.parquet", engine="pyarrow")

creates a parquet file with three columns if you use pyarrow for serialization: a, b, and __index_level_0__. If you’re using fastparquet, the index may or may not be written to the file.

This unexpected extra column causes some databases like Amazon Redshift to reject the file, because that column doesn’t exist in the target table.

If you want to omit a dataframe’s indexes when writing, pass index=False to to_parquet():

In [627]: df.to_parquet("test.parquet", index=False)

This creates a parquet file with just the two expected columns, a and b. If your DataFrame has a custom index, you won’t get it back when you load this file into a DataFrame.

Passing index=True will always write the index, even if that’s not the underlying engine’s default behavior.

Partitioning Parquet files#

Parquet supports partitioning of data based on the values of one or more columns.

In [628]: df = pd.DataFrame({"a": [0, 0, 1, 1], "b": [0, 1, 0, 1]})

In [629]: df.to_parquet(path="test", engine="pyarrow", partition_cols=["a"], compression=None)

The path specifies the parent directory to which data will be saved. The partition_cols are the column names by which the dataset will be partitioned. Columns are partitioned in the order they are given. The partition splits are determined by the unique values in the partition columns. The above example creates a partitioned dataset that may look like:

test
├── a=0
│   ├── 0bac803e32dc42ae83fddfd029cbdebc.parquet
│   └──  ...
└── a=1
    ├── e6ab24a4f45147b49b54a662f0c412a3.parquet
    └── ...

ORC#

Similar to the parquet format, the ORC Format is a binary columnar serialization for data frames. It is designed to make reading data frames efficient. pandas provides both the reader and the writer for the ORC format, read_orc() and to_orc(). This requires the pyarrow library.

Warning

In [630]: df = pd.DataFrame(
   .....:     {
   .....:         "a": list("abc"),
   .....:         "b": list(range(1, 4)),
   .....:         "c": np.arange(4.0, 7.0, dtype="float64"),
   .....:         "d": [True, False, True],
   .....:         "e": pd.date_range("20130101", periods=3),
   .....:     }
   .....: )
   .....: 

In [631]: df
Out[631]: 
   a  b    c      d          e
0  a  1  4.0   True 2013-01-01
1  b  2  5.0  False 2013-01-02
2  c  3  6.0   True 2013-01-03

In [632]: df.dtypes
Out[632]: 
a            object
b             int64
c           float64
d              bool
e    datetime64[ns]
dtype: object

Write to an orc file.

In [633]: df.to_orc("example_pa.orc", engine="pyarrow")

Read from an orc file.

In [634]: result = pd.read_orc("example_pa.orc")

In [635]: result.dtypes
Out[635]: 
a            object
b             int64
c           float64
d              bool
e    datetime64[ns]
dtype: object

Read only certain columns of an orc file.

In [636]: result = pd.read_orc(
   .....:     "example_pa.orc",
   .....:     columns=["a", "b"],
   .....: )
   .....: 

In [637]: result.dtypes
Out[637]: 
a    object
b     int64
dtype: object

SQL queries#

The pandas.io.sql module provides a collection of query wrappers to both facilitate data retrieval and to reduce dependency on DB-specific API. Database abstraction is provided by SQLAlchemy if installed. In addition you will need a driver library for your database. Examples of such drivers are psycopg2 for PostgreSQL or pymysql for MySQL. For SQLite this is included in Python’s standard library by default. You can find an overview of supported drivers for each SQL dialect in the SQLAlchemy docs.

If SQLAlchemy is not installed, you can use a sqlite3.Connection in place of a SQLAlchemy engine, connection, or URI string.

See also some cookbook examples for some advanced strategies.

The key functions are:

read_sql_table(table_name, con[, schema, ...])

Read SQL database table into a DataFrame.

read_sql_query(sql, con[, index_col, ...])

Read SQL query into a DataFrame.

read_sql(sql, con[, index_col, ...])

Read SQL query or database table into a DataFrame.

DataFrame.to_sql(name, con, *[, schema, ...])

Write records stored in a DataFrame to a SQL database.

Note

The function read_sql() is a convenience wrapper around read_sql_table() and read_sql_query() (and for backward compatibility) and will delegate to specific function depending on the provided input (database table name or sql query). Table names do not need to be quoted if they have special characters.

In the following example, we use the SQlite SQL database engine. You can use a temporary SQLite database where data are stored in “memory”.

To connect with SQLAlchemy you use the create_engine() function to create an engine object from database URI. You only need to create the engine once per database you are connecting to. For more information on create_engine() and the URI formatting, see the examples below and the SQLAlchemy documentation

In [638]: from sqlalchemy import create_engine

# Create your engine.
In [639]: engine = create_engine("sqlite:///:memory:")

If you want to manage your own connections you can pass one of those instead. The example below opens a connection to the database using a Python context manager that automatically closes the connection after the block has completed. See the SQLAlchemy docs for an explanation of how the database connection is handled.

with engine.connect() as conn, conn.begin():
    data = pd.read_sql_table("data", conn)

Warning

When you open a connection to a database you are also responsible for closing it. Side effects of leaving a connection open may include locking the database or other breaking behaviour.

Writing DataFrames#

Assuming the following data is in a DataFrame data, we can insert it into the database using to_sql().

id

Date

Col_1

Col_2

Col_3

26

2012-10-18

X

25.7

True

42

2012-10-19

Y

-12.4

False

63

2012-10-20

Z

5.73

True

In [640]: import datetime

In [641]: c = ["id", "Date", "Col_1", "Col_2", "Col_3"]

In [642]: d = [
   .....:     (26, datetime.datetime(2010, 10, 18), "X", 27.5, True),
   .....:     (42, datetime.datetime(2010, 10, 19), "Y", -12.5, False),
   .....:     (63, datetime.datetime(2010, 10, 20), "Z", 5.73, True),
   .....: ]
   .....: 

In [643]: data = pd.DataFrame(d, columns=c)

In [644]: data
Out[644]: 
   id       Date Col_1  Col_2  Col_3
0  26 2010-10-18     X  27.50   True
1  42 2010-10-19     Y -12.50  False
2  63 2010-10-20     Z   5.73   True

In [645]: data.to_sql("data", con=engine)
Out[645]: 3

With some databases, writing large DataFrames can result in errors due to packet size limitations being exceeded. This can be avoided by setting the chunksize parameter when calling to_sql. For example, the following writes data to the database in batches of 1000 rows at a time:

In [646]: data.to_sql("data_chunked", con=engine, chunksize=1000)
Out[646]: 3

SQL data types#

to_sql() will try to map your data to an appropriate SQL data type based on the dtype of the data. When you have columns of dtype object, pandas will try to infer the data type.

You can always override the default type by specifying the desired SQL type of any of the columns by using the dtype argument. This argument needs a dictionary mapping column names to SQLAlchemy types (or strings for the sqlite3 fallback mode). For example, specifying to use the sqlalchemy String type instead of the default Text type for string columns:

In [647]: from sqlalchemy.types import String

In [648]: data.to_sql("data_dtype", con=engine, dtype={"Col_1": String})
Out[648]: 3

Note

Due to the limited support for timedelta’s in the different database flavors, columns with type timedelta64 will be written as integer values as nanoseconds to the database and a warning will be raised.

Note

Columns of category dtype will be converted to the dense representation as you would get with np.asarray(categorical) (e.g. for string categories this gives an array of strings). Because of this, reading the database table back in does not generate a categorical.

Datetime data types#

Using SQLAlchemy, to_sql() is capable of writing datetime data that is timezone naive or timezone aware. However, the resulting data stored in the database ultimately depends on the supported data type for datetime data of the database system being used.

The following table lists supported data types for datetime data for some common databases. Other database dialects may have different data types for datetime data.

Database

SQL Datetime Types

Timezone Support

SQLite

TEXT

No

MySQL

TIMESTAMP or DATETIME

No

PostgreSQL

TIMESTAMP or TIMESTAMP WITH TIME ZONE

Yes

When writing timezone aware data to databases that do not support timezones, the data will be written as timezone naive timestamps that are in local time with respect to the timezone.

read_sql_table() is also capable of reading datetime data that is timezone aware or naive. When reading TIMESTAMP WITH TIME ZONE types, pandas will convert the data to UTC.

Insertion method#

The parameter method controls the SQL insertion clause used. Possible values are:

  • None: Uses standard SQL INSERT clause (one per row).

  • 'multi': Pass multiple values in a single INSERT clause. It uses a special SQL syntax not supported by all backends. This usually provides better performance for analytic databases like Presto and Redshift, but has worse performance for traditional SQL backend if the table contains many columns. For more information check the SQLAlchemy documentation.

  • callable with signature (pd_table, conn, keys, data_iter): This can be used to implement a more performant insertion method based on specific backend dialect features.

Example of a callable using PostgreSQL COPY clause:

# Alternative to_sql() *method* for DBs that support COPY FROM
import csv
from io import StringIO

def psql_insert_copy(table, conn, keys, data_iter):
    """
    Execute SQL statement inserting data

    Parameters
    ----------
    table : pandas.io.sql.SQLTable
    conn : sqlalchemy.engine.Engine or sqlalchemy.engine.Connection
    keys : list of str
        Column names
    data_iter : Iterable that iterates the values to be inserted
    """
    # gets a DBAPI connection that can provide a cursor
    dbapi_conn = conn.connection
    with dbapi_conn.cursor() as cur:
        s_buf = StringIO()
        writer = csv.writer(s_buf)
        writer.writerows(data_iter)
        s_buf.seek(0)

        columns = ', '.join(['"{}"'.format(k) for k in keys])
        if table.schema:
            table_name = '{}.{}'.format(table.schema, table.name)
        else:
            table_name = table.name

        sql = 'COPY {} ({}) FROM STDIN WITH CSV'.format(
            table_name, columns)
        cur.copy_expert(sql=sql, file=s_buf)

Reading tables#

read_sql_table() will read a database table given the table name and optionally a subset of columns to read.

Note

In order to use read_sql_table(), you must have the SQLAlchemy optional dependency installed.

In [649]: pd.read_sql_table("data", engine)
Out[649]: 
   index  id       Date Col_1  Col_2  Col_3
0      0  26 2010-10-18     X  27.50   True
1      1  42 2010-10-19     Y -12.50  False
2      2  63 2010-10-20     Z   5.73   True

Note

Note that pandas infers column dtypes from query outputs, and not by looking up data types in the physical database schema. For example, assume userid is an integer column in a table. Then, intuitively, select userid ... will return integer-valued series, while select cast(userid as text) ... will return object-valued (str) series. Accordingly, if the query output is empty, then all resulting columns will be returned as object-valued (since they are most general). If you foresee that your query will sometimes generate an empty result, you may want to explicitly typecast afterwards to ensure dtype integrity.

You can also specify the name of the column as the DataFrame index, and specify a subset of columns to be read.

In [650]: pd.read_sql_table("data", engine, index_col="id")
Out[650]: 
    index       Date Col_1  Col_2  Col_3
id                                      
26      0 2010-10-18     X  27.50   True
42      1 2010-10-19     Y -12.50  False
63      2 2010-10-20     Z   5.73   True

In [651]: pd.read_sql_table("data", engine, columns=["Col_1", "Col_2"])
Out[651]: 
  Col_1  Col_2
0     X  27.50
1     Y -12.50
2     Z   5.73

And you can explicitly force columns to be parsed as dates:

In [652]: pd.read_sql_table("data", engine, parse_dates=["Date"])
Out[652]: 
   index  id       Date Col_1  Col_2  Col_3
0      0  26 2010-10-18     X  27.50   True
1      1  42 2010-10-19     Y -12.50  False
2      2  63 2010-10-20     Z   5.73   True

If needed you can explicitly specify a format string, or a dict of arguments to pass to pandas.to_datetime():

pd.read_sql_table("data", engine, parse_dates={"Date": "%Y-%m-%d"})
pd.read_sql_table(
    "data",
    engine,
    parse_dates={"Date": {"format": "%Y-%m-%d %H:%M:%S"}},
)

You can check if a table exists using has_table()

Schema support#

Reading from and writing to different schema’s is supported through the schema keyword in the read_sql_table() and to_sql() functions. Note however that this depends on the database flavor (sqlite does not have schema’s). For example:

df.to_sql(name="table", con=engine, schema="other_schema")
pd.read_sql_table("table", engine, schema="other_schema")

Querying#

You can query using raw SQL in the read_sql_query() function. In this case you must use the SQL variant appropriate for your database. When using SQLAlchemy, you can also pass SQLAlchemy Expression language constructs, which are database-agnostic.

In [653]: pd.read_sql_query("SELECT * FROM data", engine)
Out[653]: 
   index  id                        Date Col_1  Col_2  Col_3
0      0  26  2010-10-18 00:00:00.000000     X  27.50      1
1      1  42  2010-10-19 00:00:00.000000     Y -12.50      0
2      2  63  2010-10-20 00:00:00.000000     Z   5.73      1

Of course, you can specify a more “complex” query.

In [654]: pd.read_sql_query("SELECT id, Col_1, Col_2 FROM data WHERE id = 42;", engine)
Out[654]: 
   id Col_1  Col_2
0  42     Y  -12.5

The read_sql_query() function supports a chunksize argument. Specifying this will return an iterator through chunks of the query result:

In [655]: df = pd.DataFrame(np.random.randn(20, 3), columns=list("abc"))

In [656]: df.to_sql(name="data_chunks", con=engine, index=False)
Out[656]: 20
In [657]: for chunk in pd.read_sql_query("SELECT * FROM data_chunks", engine, chunksize=5):
   .....:     print(chunk)
   .....: 
          a         b         c
0 -0.395347 -0.822726 -0.363777
1  1.676124 -0.908102 -1.391346
2 -1.094269  0.278380  1.205899
3  1.503443  0.932171 -0.709459
4 -0.645944 -1.351389  0.132023
          a         b         c
0  0.210427  0.192202  0.661949
1  1.690629 -1.046044  0.618697
2 -0.013863  1.314289  1.951611
3 -1.485026  0.304662  1.194757
4 -0.446717  0.528496 -0.657575
          a         b         c
0 -0.876654  0.336252  0.172668
1  0.337684 -0.411202 -0.828394
2 -0.244413  1.094948  0.087183
3  1.125934 -1.480095  1.205944
4 -0.451849  0.452214 -2.208192
          a         b         c
0 -2.061019  0.044184 -0.017118
1  1.248959 -0.675595 -1.908296
2 -0.125934  1.491974  0.648726
3  0.391214  0.438609  1.634248
4  1.208707 -1.535740  1.620399

Engine connection examples#

To connect with SQLAlchemy you use the create_engine() function to create an engine object from database URI. You only need to create the engine once per database you are connecting to.

from sqlalchemy import create_engine

engine = create_engine("postgresql://scott:tiger@localhost:5432/mydatabase")

engine = create_engine("mysql+mysqldb://scott:tiger@localhost/foo")

engine = create_engine("oracle://scott:[email protected]:1521/sidname")

engine = create_engine("mssql+pyodbc://mydsn")

# sqlite://<nohostname>/<path>
# where <path> is relative:
engine = create_engine("sqlite:///foo.db")

# or absolute, starting with a slash:
engine = create_engine("sqlite:////absolute/path/to/foo.db")

For more information see the examples the SQLAlchemy documentation

Advanced SQLAlchemy queries#

You can use SQLAlchemy constructs to describe your query.

Use sqlalchemy.text() to specify query parameters in a backend-neutral way

In [658]: import sqlalchemy as sa

In [659]: pd.read_sql(
   .....:     sa.text("SELECT * FROM data where Col_1=:col1"), engine, params={"col1": "X"}
   .....: )
   .....: 
Out[659]: 
   index  id                        Date Col_1  Col_2  Col_3
0      0  26  2010-10-18 00:00:00.000000     X   27.5      1

If you have an SQLAlchemy description of your database you can express where conditions using SQLAlchemy expressions

In [660]: metadata = sa.MetaData()

In [661]: data_table = sa.Table(
   .....:     "data",
   .....:     metadata,
   .....:     sa.Column("index", sa.Integer),
   .....:     sa.Column("Date", sa.DateTime),
   .....:     sa.Column("Col_1", sa.String),
   .....:     sa.Column("Col_2", sa.Float),
   .....:     sa.Column("Col_3", sa.Boolean),
   .....: )
   .....: 

In [662]: pd.read_sql(sa.select(data_table).where(data_table.c.Col_3 is True), engine)
Out[662]: 
Empty DataFrame
Columns: [index, Date, Col_1, Col_2, Col_3]
Index: []

You can combine SQLAlchemy expressions with parameters passed to read_sql() using sqlalchemy.bindparam()

In [663]: import datetime as dt

In [664]: expr = sa.select(data_table).where(data_table.c.Date > sa.bindparam("date"))

In [665]: pd.read_sql(expr, engine, params={"date": dt.datetime(2010, 10, 18)})
Out[665]: 
   index       Date Col_1  Col_2  Col_3
0      1 2010-10-19     Y -12.50  False
1      2 2010-10-20     Z   5.73   True

Sqlite fallback#

The use of sqlite is supported without using SQLAlchemy. This mode requires a Python database adapter which respect the Python DB-API.

You can create connections like so:

import sqlite3

con = sqlite3.connect(":memory:")

And then issue the following queries:

data.to_sql("data", con)
pd.read_sql_query("SELECT * FROM data", con)

Google BigQuery#

The pandas-gbq package provides functionality to read/write from Google BigQuery.

pandas integrates with this external package. if pandas-gbq is installed, you can use the pandas methods pd.read_gbq and DataFrame.to_gbq, which will call the respective functions from pandas-gbq.

Full documentation can be found here.

Stata format#

Writing to stata format#

The method to_stata() will write a DataFrame into a .dta file. The format version of this file is always 115 (Stata 12).

In [666]: df = pd.DataFrame(np.random.randn(10, 2), columns=list("AB"))

In [667]: df.to_stata("stata.dta")

Stata data files have limited data type support; only strings with 244 or fewer characters, int8, int16, int32, float32 and float64 can be stored in .dta files. Additionally, Stata reserves certain values to represent missing data. Exporting a non-missing value that is outside of the permitted range in Stata for a particular data type will retype the variable to the next larger size. For example, int8 values are restricted to lie between -127 and 100 in Stata, and so variables with values above 100 will trigger a conversion to int16. nan values in floating points data types are stored as the basic missing data type (. in Stata).

Note

It is not possible to export missing data values for integer data types.

The Stata writer gracefully handles other data types including int64, bool, uint8, uint16, uint32 by casting to the smallest supported type that can represent the data. For example, data with a type of uint8 will be cast to int8 if all values are less than 100 (the upper bound for non-missing int8 data in Stata), or, if values are outside of this range, the variable is cast to int16.

Warning

Conversion from int64 to float64 may result in a loss of precision if int64 values are larger than 2**53.

Warning

StataWriter and to_stata() only support fixed width strings containing up to 244 characters, a limitation imposed by the version 115 dta file format. Attempting to write Stata dta files with strings longer than 244 characters raises a ValueError.

Reading from Stata format#

The top-level function read_stata will read a dta file and return either a DataFrame or a pandas.api.typing.StataReader that can be used to read the file incrementally.

In [668]: pd.read_stata("stata.dta")
Out[668]: 
   index         A         B
0      0 -0.165614  0.490482
1      1 -0.637829  0.067091
2      2 -0.242577  1.348038
3      3  0.647699 -0.644937
4      4  0.625771  0.918376
5      5  0.401781 -1.488919
6      6 -0.981845 -0.046882
7      7 -0.306796  0.877025
8      8 -0.336606  0.624747
9      9 -1.582600  0.806340

Specifying a chunksize yields a pandas.api.typing.StataReader instance that can be used to read chunksize lines from the file at a time. The StataReader object can be used as an iterator.

In [669]: with pd.read_stata("stata.dta", chunksize=3) as reader:
   .....:     for df in reader:
   .....:         print(df.shape)
   .....: 
(3, 3)
(3, 3)
(3, 3)
(1, 3)

For more fine-grained control, use iterator=True and specify chunksize with each call to read().

In [670]: with pd.read_stata("stata.dta", iterator=True) as reader:
   .....:     chunk1 = reader.read(5)
   .....:     chunk2 = reader.read(5)
   .....: 

Currently the index is retrieved as a column.

The parameter convert_categoricals indicates whether value labels should be read and used to create a Categorical variable from them. Value labels can also be retrieved by the function value_labels, which requires read() to be called before use.

The parameter convert_missing indicates whether missing value representations in Stata should be preserved. If False (the default), missing values are represented as np.nan. If True, missing values are represented using StataMissingValue objects, and columns containing missing values will have object data type.

Note

read_stata() and StataReader support .dta formats 113-115 (Stata 10-12), 117 (Stata 13), and 118 (Stata 14).

Note

Setting preserve_dtypes=False will upcast to the standard pandas data types: int64 for all integer types and float64 for floating point data. By default, the Stata data types are preserved when importing.

Note

All StataReader objects, whether created by read_stata() (when using iterator=True or chunksize) or instantiated by hand, must be used as context managers (e.g. the with statement). While the close() method is available, its use is unsupported. It is not part of the public API and will be removed in with future without warning.

Categorical data#

Categorical data can be exported to Stata data files as value labeled data. The exported data consists of the underlying category codes as integer data values and the categories as value labels. Stata does not have an explicit equivalent to a Categorical and information about whether the variable is ordered is lost when exporting.

Warning

Stata only supports string value labels, and so str is called on the categories when exporting data. Exporting Categorical variables with non-string categories produces a warning, and can result a loss of information if the str representations of the categories are not unique.

Labeled data can similarly be imported from Stata data files as Categorical variables using the keyword argument convert_categoricals (True by default). The keyword argument order_categoricals (True by default) determines whether imported Categorical variables are ordered.

Note

When importing categorical data, the values of the variables in the Stata data file are not preserved since Categorical variables always use integer data types between -1 and n-1 where n is the number of categories. If the original values in the Stata data file are required, these can be imported by setting convert_categoricals=False, which will import original data (but not the variable labels). The original values can be matched to the imported categorical data since there is a simple mapping between the original Stata data values and the category codes of imported Categorical variables: missing values are assigned code -1, and the smallest original value is assigned 0, the second smallest is assigned 1 and so on until the largest original value is assigned the code n-1.

Note

Stata supports partially labeled series. These series have value labels for some but not all data values. Importing a partially labeled series will produce a Categorical with string categories for the values that are labeled and numeric categories for values with no label.

SAS formats#

The top-level function read_sas() can read (but not write) SAS XPORT (.xpt) and SAS7BDAT (.sas7bdat) format files.

SAS files only contain two value types: ASCII text and floating point values (usually 8 bytes but sometimes truncated). For xport files, there is no automatic type conversion to integers, dates, or categoricals. For SAS7BDAT files, the format codes may allow date variables to be automatically converted to dates. By default the whole file is read and returned as a DataFrame.

Specify a chunksize or use iterator=True to obtain reader objects (XportReader or SAS7BDATReader) for incrementally reading the file. The reader objects also have attributes that contain additional information about the file and its variables.

Read a SAS7BDAT file:

df = pd.read_sas("sas_data.sas7bdat")

Obtain an iterator and read an XPORT file 100,000 lines at a time:

def do_something(chunk):
    pass


with pd.read_sas("sas_xport.xpt", chunk=100000) as rdr:
    for chunk in rdr:
        do_something(chunk)

The specification for the xport file format is available from the SAS web site.

No official documentation is available for the SAS7BDAT format.

SPSS formats#

The top-level function read_spss() can read (but not write) SPSS SAV (.sav) and ZSAV (.zsav) format files.

SPSS files contain column names. By default the whole file is read, categorical columns are converted into pd.Categorical, and a DataFrame with all columns is returned.

Specify the usecols parameter to obtain a subset of columns. Specify convert_categoricals=False to avoid converting categorical columns into pd.Categorical.

Read an SPSS file:

df = pd.read_spss("spss_data.sav")

Extract a subset of columns contained in usecols from an SPSS file and avoid converting categorical columns into pd.Categorical:

df = pd.read_spss(
    "spss_data.sav",
    usecols=["foo", "bar"],
    convert_categoricals=False,
)

More information about the SAV and ZSAV file formats is available here.

Other file formats#

pandas itself only supports IO with a limited set of file formats that map cleanly to its tabular data model. For reading and writing other file formats into and from pandas, we recommend these packages from the broader community.

netCDF#

xarray provides data structures inspired by the pandas DataFrame for working with multi-dimensional datasets, with a focus on the netCDF file format and easy conversion to and from pandas.

Performance considerations#

This is an informal comparison of various IO methods, using pandas 0.24.2. Timings are machine dependent and small differences should be ignored.

In [1]: sz = 1000000
In [2]: df = pd.DataFrame({'A': np.random.randn(sz), 'B': [1] * sz})

In [3]: df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 1000000 entries, 0 to 999999
Data columns (total 2 columns):
A    1000000 non-null float64
B    1000000 non-null int64
dtypes: float64(1), int64(1)
memory usage: 15.3 MB

The following test functions will be used below to compare the performance of several IO methods:

import numpy as np

import os

sz = 1000000
df = pd.DataFrame({"A": np.random.randn(sz), "B": [1] * sz})

sz = 1000000
np.random.seed(42)
df = pd.DataFrame({"A": np.random.randn(sz), "B": [1] * sz})


def test_sql_write(df):
    if os.path.exists("test.sql"):
        os.remove("test.sql")
    sql_db = sqlite3.connect("test.sql")
    df.to_sql(name="test_table", con=sql_db)
    sql_db.close()


def test_sql_read():
    sql_db = sqlite3.connect("test.sql")
    pd.read_sql_query("select * from test_table", sql_db)
    sql_db.close()


def test_hdf_fixed_write(df):
    df.to_hdf("test_fixed.hdf", "test", mode="w")


def test_hdf_fixed_read():
    pd.read_hdf("test_fixed.hdf", "test")


def test_hdf_fixed_write_compress(df):
    df.to_hdf("test_fixed_compress.hdf", "test", mode="w", complib="blosc")


def test_hdf_fixed_read_compress():
    pd.read_hdf("test_fixed_compress.hdf", "test")


def test_hdf_table_write(df):
    df.to_hdf("test_table.hdf", "test", mode="w", format="table")


def test_hdf_table_read():
    pd.read_hdf("test_table.hdf", "test")


def test_hdf_table_write_compress(df):
    df.to_hdf(
        "test_table_compress.hdf", "test", mode="w", complib="blosc", format="table"
    )


def test_hdf_table_read_compress():
    pd.read_hdf("test_table_compress.hdf", "test")


def test_csv_write(df):
    df.to_csv("test.csv", mode="w")


def test_csv_read():
    pd.read_csv("test.csv", index_col=0)


def test_feather_write(df):
    df.to_feather("test.feather")


def test_feather_read():
    pd.read_feather("test.feather")


def test_pickle_write(df):
    df.to_pickle("test.pkl")


def test_pickle_read():
    pd.read_pickle("test.pkl")


def test_pickle_write_compress(df):
    df.to_pickle("test.pkl.compress", compression="xz")


def test_pickle_read_compress():
    pd.read_pickle("test.pkl.compress", compression="xz")


def test_parquet_write(df):
    df.to_parquet("test.parquet")


def test_parquet_read():
    pd.read_parquet("test.parquet")

When writing, the top three functions in terms of speed are test_feather_write, test_hdf_fixed_write and test_hdf_fixed_write_compress.

In [4]: %timeit test_sql_write(df)
3.29 s ± 43.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [5]: %timeit test_hdf_fixed_write(df)
19.4 ms ± 560 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [6]: %timeit test_hdf_fixed_write_compress(df)
19.6 ms ± 308 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [7]: %timeit test_hdf_table_write(df)
449 ms ± 5.61 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [8]: %timeit test_hdf_table_write_compress(df)
448 ms ± 11.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [9]: %timeit test_csv_write(df)
3.66 s ± 26.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [10]: %timeit test_feather_write(df)
9.75 ms ± 117 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [11]: %timeit test_pickle_write(df)
30.1 ms ± 229 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [12]: %timeit test_pickle_write_compress(df)
4.29 s ± 15.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [13]: %timeit test_parquet_write(df)
67.6 ms ± 706 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

When reading, the top three functions in terms of speed are test_feather_read, test_pickle_read and test_hdf_fixed_read.

In [14]: %timeit test_sql_read()
1.77 s ± 17.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [15]: %timeit test_hdf_fixed_read()
19.4 ms ± 436 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [16]: %timeit test_hdf_fixed_read_compress()
19.5 ms ± 222 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [17]: %timeit test_hdf_table_read()
38.6 ms ± 857 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [18]: %timeit test_hdf_table_read_compress()
38.8 ms ± 1.49 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [19]: %timeit test_csv_read()
452 ms ± 9.04 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [20]: %timeit test_feather_read()
12.4 ms ± 99.7 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [21]: %timeit test_pickle_read()
18.4 ms ± 191 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [22]: %timeit test_pickle_read_compress()
915 ms ± 7.48 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [23]: %timeit test_parquet_read()
24.4 ms ± 146 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

The files test.pkl.compress, test.parquet and test.feather took the least space on disk (in bytes).

29519500 Oct 10 06:45 test.csv
16000248 Oct 10 06:45 test.feather
8281983  Oct 10 06:49 test.parquet
16000857 Oct 10 06:47 test.pkl
7552144  Oct 10 06:48 test.pkl.compress
34816000 Oct 10 06:42 test.sql
24009288 Oct 10 06:43 test_fixed.hdf
24009288 Oct 10 06:43 test_fixed_compress.hdf
24458940 Oct 10 06:44 test_table.hdf
24458940 Oct 10 06:44 test_table_compress.hdf