v0.21.0 (October 27, 2017)

This is a major release from 0.20.3 and includes a number of API changes, deprecations, new features, enhancements, and performance improvements along with a large number of bug fixes. We recommend that all users upgrade to this version.

Highlights include:

  • Integration with Apache Parquet, including a new top-level read_parquet() function and DataFrame.to_parquet() method, see here.
  • New user-facing pandas.api.types.CategoricalDtype for specifying categoricals independent of the data, see here.
  • The behavior of sum and prod on all-NaN Series/DataFrames is now consistent and no longer depends on whether bottleneck is installed, and sum and prod on empty Series now return NaN instead of 0, see here.
  • Compatibility fixes for pypy, see here.
  • Additions to the drop, reindex and rename API to make them more consistent, see here.
  • Addition of the new methods DataFrame.infer_objects (see here) and GroupBy.pipe (see here).
  • Indexing with a list of labels, where one or more of the labels is missing, is deprecated and will raise a KeyError in a future version, see here.

Check the API Changes and deprecations before updating.

New features

Integration with Apache Parquet file format

Integration with Apache Parquet, including a new top-level read_parquet() and DataFrame.to_parquet() method, see here (GH15838, GH17438).

Apache Parquet provides a cross-language, binary file format for reading and writing data frames efficiently. Parquet is designed to faithfully serialize and de-serialize DataFrame s, supporting all of the pandas dtypes, including extension dtypes such as datetime with timezones.

This functionality depends on either the pyarrow or fastparquet library. For more details, see see the IO docs on Parquet.

infer_objects type conversion

The DataFrame.infer_objects() and Series.infer_objects() methods have been added to perform dtype inference on object columns, replacing some of the functionality of the deprecated convert_objects method. See the documentation here for more details. (GH11221)

This method only performs soft conversions on object columns, converting Python objects to native types, but not any coercive conversions. For example:

In [1]: df = pd.DataFrame({'A': [1, 2, 3],
   ...:                    'B': np.array([1, 2, 3], dtype='object'),
   ...:                    'C': ['1', '2', '3']})

In [2]: df.dtypes
A     int64
B    object
C    object
Length: 3, dtype: object

In [3]: df.infer_objects().dtypes
A     int64
B     int64
C    object
Length: 3, dtype: object

Note that column 'C' was not converted - only scalar numeric types will be converted to a new type. Other types of conversion should be accomplished using the to_numeric() function (or to_datetime(), to_timedelta()).

In [4]: df = df.infer_objects()

In [5]: df['C'] = pd.to_numeric(df['C'], errors='coerce')

In [6]: df.dtypes
A    int64
B    int64
C    int64
Length: 3, dtype: object

Improved warnings when attempting to create columns

New users are often puzzled by the relationship between column operations and attribute access on DataFrame instances (GH7175). One specific instance of this confusion is attempting to create a new column by setting an attribute on the DataFrame:

In [1]: df = pd.DataFrame({'one': [1., 2., 3.]})
In [2]: df.two = [4, 5, 6]

This does not raise any obvious exceptions, but also does not create a new column:

In [3]: df
0  1.0
1  2.0
2  3.0

Setting a list-like data structure into a new attribute now raises a UserWarning about the potential for unexpected behavior. See Attribute Access.

drop now also accepts index/columns keywords

The drop() method has gained index/columns keywords as an alternative to specifying the axis. This is similar to the behavior of reindex (GH12392).

For example:

In [7]: df = pd.DataFrame(np.arange(8).reshape(2, 4),
   ...:                   columns=['A', 'B', 'C', 'D'])

In [8]: df
   A  B  C  D
0  0  1  2  3
1  4  5  6  7

[2 rows x 4 columns]

In [9]: df.drop(['B', 'C'], axis=1)
   A  D
0  0  3
1  4  7

[2 rows x 2 columns]

# the following is now equivalent
In [10]: df.drop(columns=['B', 'C'])
   A  D
0  0  3
1  4  7

[2 rows x 2 columns]

rename, reindex now also accept axis keyword

The DataFrame.rename() and DataFrame.reindex() methods have gained the axis keyword to specify the axis to target with the operation (GH12392).

Here’s rename:

In [11]: df = pd.DataFrame({"A": [1, 2, 3], "B": [4, 5, 6]})

In [12]: df.rename(str.lower, axis='columns')
   a  b
0  1  4
1  2  5
2  3  6

[3 rows x 2 columns]

In [13]: df.rename(id, axis='index')
                A  B
94683363768992  1  4
94683363769024  2  5
94683363769056  3  6

[3 rows x 2 columns]

And reindex:

In [14]: df.reindex(['A', 'B', 'C'], axis='columns')
   A  B   C
0  1  4 NaN
1  2  5 NaN
2  3  6 NaN

[3 rows x 3 columns]

In [15]: df.reindex([0, 1, 3], axis='index')
     A    B
0  1.0  4.0
1  2.0  5.0
3  NaN  NaN

[3 rows x 2 columns]

The “index, columns” style continues to work as before.

In [16]: df.rename(index=id, columns=str.lower)
                a  b
94683363768992  1  4
94683363769024  2  5
94683363769056  3  6

[3 rows x 2 columns]

In [17]: df.reindex(index=[0, 1, 3], columns=['A', 'B', 'C'])
     A    B   C
0  1.0  4.0 NaN
1  2.0  5.0 NaN
3  NaN  NaN NaN

[3 rows x 3 columns]

We highly encourage using named arguments to avoid confusion when using either style.

CategoricalDtype for specifying categoricals

pandas.api.types.CategoricalDtype has been added to the public API and expanded to include the categories and ordered attributes. A CategoricalDtype can be used to specify the set of categories and orderedness of an array, independent of the data. This can be useful for example, when converting string data to a Categorical (GH14711, GH15078, GH16015, GH17643):

In [18]: from pandas.api.types import CategoricalDtype

In [19]: s = pd.Series(['a', 'b', 'c', 'a'])  # strings

In [20]: dtype = CategoricalDtype(categories=['a', 'b', 'c', 'd'], ordered=True)

In [21]: s.astype(dtype)
0    a
1    b
2    c
3    a
Length: 4, dtype: category
Categories (4, object): [a < b < c < d]

One place that deserves special mention is in read_csv(). Previously, with dtype={'col': 'category'}, the returned values and categories would always be strings.

In [22]: data = 'A,B\na,1\nb,2\nc,3'

In [23]: pd.read_csv(StringIO(data), dtype={'B': 'category'}).B.cat.categories
Out[23]: Index(['1', '2', '3'], dtype='object')

Notice the “object” dtype.

With a CategoricalDtype of all numerics, datetimes, or timedeltas, we can automatically convert to the correct type

In [24]: dtype = {'B': CategoricalDtype([1, 2, 3])}

In [25]: pd.read_csv(StringIO(data), dtype=dtype).B.cat.categories
Out[25]: Int64Index([1, 2, 3], dtype='int64')

The values have been correctly interpreted as integers.

The .dtype property of a Categorical, CategoricalIndex or a Series with categorical type will now return an instance of CategoricalDtype. While the repr has changed, str(CategoricalDtype()) is still the string 'category'. We’ll take this moment to remind users that the preferred way to detect categorical data is to use pandas.api.types.is_categorical_dtype(), and not str(dtype) == 'category'.

See the CategoricalDtype docs for more.

GroupBy objects now have a pipe method

GroupBy objects now have a pipe method, similar to the one on DataFrame and Series, that allow for functions that take a GroupBy to be composed in a clean, readable syntax. (GH17871)

For a concrete example on combining .groupby and .pipe , imagine having a DataFrame with columns for stores, products, revenue and sold quantity. We’d like to do a groupwise calculation of prices (i.e. revenue/quantity) per store and per product. We could do this in a multi-step operation, but expressing it in terms of piping can make the code more readable.

First we set the data:

In [26]: import numpy as np

In [27]: n = 1000

In [28]: df = pd.DataFrame({'Store': np.random.choice(['Store_1', 'Store_2'], n),
   ....:                    'Product': np.random.choice(['Product_1',
   ....:                                                 'Product_2',
   ....:                                                 'Product_3'
   ....:                                                 ], n),
   ....:                    'Revenue': (np.random.random(n) * 50 + 10).round(2),
   ....:                    'Quantity': np.random.randint(1, 10, size=n)})

In [29]: df.head(2)
     Store    Product  Revenue  Quantity
0  Store_2  Product_2    32.09         7
1  Store_1  Product_3    14.20         1

[2 rows x 4 columns]

Now, to find prices per store/product, we can simply do:

In [30]: (df.groupby(['Store', 'Product'])
   ....:    .pipe(lambda grp: grp.Revenue.sum() / grp.Quantity.sum())
   ....:    .unstack().round(2))
Product  Product_1  Product_2  Product_3
Store_1       6.73       6.72       7.14
Store_2       7.59       6.98       7.23

[2 rows x 3 columns]

See the documentation for more.

Categorical.rename_categories accepts a dict-like

rename_categories() now accepts a dict-like argument for new_categories. The previous categories are looked up in the dictionary’s keys and replaced if found. The behavior of missing and extra keys is the same as in DataFrame.rename().

In [31]: c = pd.Categorical(['a', 'a', 'b'])

In [32]: c.rename_categories({"a": "eh", "b": "bee"})
[eh, eh, bee]
Categories (2, object): [eh, bee]


To assist with upgrading pandas, rename_categories treats Series as list-like. Typically, Series are considered to be dict-like (e.g. in .rename, .map). In a future version of pandas rename_categories will change to treat them as dict-like. Follow the warning message’s recommendations for writing future-proof code.

In [33]: c.rename_categories(pd.Series([0, 1], index=['a', 'c']))
FutureWarning: Treating Series 'new_categories' as a list-like and using the values.
In a future version, 'rename_categories' will treat Series like a dictionary.
For dict-like, use 'new_categories.to_dict()'
For list-like, use 'new_categories.values'.
[0, 0, 1]
Categories (2, int64): [0, 1]

Other Enhancements

New functions or methods

New keywords

Various enhancements

Backwards incompatible API changes

Dependencies have increased minimum versions

We have updated our minimum supported versions of dependencies (GH15206, GH15543, GH15214). If installed, we now require:

Package Minimum Version Required
Numpy 1.9.0 X
Matplotlib 1.4.3  
Scipy 0.14.0  
Bottleneck 1.0.0  

Additionally, support has been dropped for Python 3.4 (GH15251).

Sum/Prod of all-NaN or empty Series/DataFrames is now consistently NaN


The changes described here have been partially reverted. See the v0.22.0 Whatsnew for more.

The behavior of sum and prod on all-NaN Series/DataFrames no longer depends on whether bottleneck is installed, and return value of sum and prod on an empty Series has changed (GH9422, GH15507).

Calling sum or prod on an empty or all-NaN Series, or columns of a DataFrame, will result in NaN. See the docs.

In [33]: s = pd.Series([np.nan])

Previously WITHOUT bottleneck installed:

In [2]: s.sum()
Out[2]: np.nan

Previously WITH bottleneck:

In [2]: s.sum()
Out[2]: 0.0

New Behavior, without regard to the bottleneck installation:

In [34]: s.sum()
Out[34]: 0.0

Note that this also changes the sum of an empty Series. Previously this always returned 0 regardless of a bottlenck installation:

In [1]: pd.Series([]).sum()
Out[1]: 0

but for consistency with the all-NaN case, this was changed to return NaN as well:

In [35]: pd.Series([]).sum()
Out[35]: 0.0

Indexing with a list with missing labels is Deprecated

Previously, selecting with a list of labels, where one or more labels were missing would always succeed, returning NaN for missing labels. This will now show a FutureWarning. In the future this will raise a KeyError (GH15747). This warning will trigger on a DataFrame or a Series for using .loc[] or [[]] when passing a list-of-labels with at least 1 missing label. See the deprecation docs.

In [36]: s = pd.Series([1, 2, 3])

In [37]: s
0    1
1    2
2    3
Length: 3, dtype: int64

Previous Behavior

In [4]: s.loc[[1, 2, 3]]
1    2.0
2    3.0
3    NaN
dtype: float64

Current Behavior

In [4]: s.loc[[1, 2, 3]]
Passing list-likes to .loc or [] with any missing label will raise
KeyError in the future, you can use .reindex() as an alternative.

See the documentation here:

1    2.0
2    3.0
3    NaN
dtype: float64

The idiomatic way to achieve selecting potentially not-found elements is via .reindex()

In [38]: s.reindex([1, 2, 3])
1    2.0
2    3.0
3    NaN
Length: 3, dtype: float64

Selection with all keys found is unchanged.

In [39]: s.loc[[1, 2]]
1    2
2    3
Length: 2, dtype: int64

NA naming Changes

In order to promote more consistency among the pandas API, we have added additional top-level functions isna() and notna() that are aliases for isnull() and notnull(). The naming scheme is now more consistent with methods like .dropna() and .fillna(). Furthermore in all cases where .isnull() and .notnull() methods are defined, these have additional methods named .isna() and .notna(), these are included for classes Categorical, Index, Series, and DataFrame. (GH15001).

The configuration option pd.options.mode.use_inf_as_null is deprecated, and pd.options.mode.use_inf_as_na is added as a replacement.

Iteration of Series/Index will now return Python scalars

Previously, when using certain iteration methods for a Series with dtype int or float, you would receive a numpy scalar, e.g. a np.int64, rather than a Python int. Issue (GH10904) corrected this for Series.tolist() and list(Series). This change makes all iteration methods consistent, in particular, for __iter__() and .map(); note that this only affects int/float dtypes. (GH13236, GH13258, GH14216).

In [40]: s = pd.Series([1, 2, 3])

In [41]: s
0    1
1    2
2    3
Length: 3, dtype: int64


In [2]: type(list(s)[0])
Out[2]: numpy.int64

New Behaviour:

In [42]: type(list(s)[0])
Out[42]: int

Furthermore this will now correctly box the results of iteration for DataFrame.to_dict() as well.

In [43]: d = {'a': [1], 'b': ['b']}

In [44]: df = pd.DataFrame(d)


In [8]: type(df.to_dict()['a'][0])
Out[8]: numpy.int64

New Behaviour:

In [45]: type(df.to_dict()['a'][0])
Out[45]: int

Indexing with a Boolean Index

Previously when passing a boolean Index to .loc, if the index of the Series/DataFrame had boolean labels, you would get a label based selection, potentially duplicating result labels, rather than a boolean indexing selection (where True selects elements), this was inconsistent how a boolean numpy array indexed. The new behavior is to act like a boolean numpy array indexer. (GH17738)

Previous Behavior:

In [46]: s = pd.Series([1, 2, 3], index=[False, True, False])

In [47]: s
False    1
True     2
False    3
Length: 3, dtype: int64
In [59]: s.loc[pd.Index([True, False, True])]
True     2
False    1
False    3
True     2
dtype: int64

Current Behavior

In [48]: s.loc[pd.Index([True, False, True])]
False    1
False    3
Length: 2, dtype: int64

Furthermore, previously if you had an index that was non-numeric (e.g. strings), then a boolean Index would raise a KeyError. This will now be treated as a boolean indexer.

Previously Behavior:

In [49]: s = pd.Series([1, 2, 3], index=['a', 'b', 'c'])

In [50]: s
a    1
b    2
c    3
Length: 3, dtype: int64
In [39]: s.loc[pd.Index([True, False, True])]
KeyError: "None of [Index([True, False, True], dtype='object')] are in the [index]"

Current Behavior

In [51]: s.loc[pd.Index([True, False, True])]
a    1
c    3
Length: 2, dtype: int64

PeriodIndex resampling

In previous versions of pandas, resampling a Series/DataFrame indexed by a PeriodIndex returned a DatetimeIndex in some cases (GH12884). Resampling to a multiplied frequency now returns a PeriodIndex (GH15944). As a minor enhancement, resampling a PeriodIndex can now handle NaT values (GH13224)

Previous Behavior:

In [1]: pi = pd.period_range('2017-01', periods=12, freq='M')

In [2]: s = pd.Series(np.arange(12), index=pi)

In [3]: resampled = s.resample('2Q').mean()

In [4]: resampled
2017-03-31     1.0
2017-09-30     5.5
2018-03-31    10.0
Freq: 2Q-DEC, dtype: float64

In [5]: resampled.index
Out[5]: DatetimeIndex(['2017-03-31', '2017-09-30', '2018-03-31'], dtype='datetime64[ns]', freq='2Q-DEC')

New Behavior:

In [52]: pi = pd.period_range('2017-01', periods=12, freq='M')

In [53]: s = pd.Series(np.arange(12), index=pi)

In [54]: resampled = s.resample('2Q').mean()

In [55]: resampled
2017Q1    2.5
2017Q3    8.5
Freq: 2Q-DEC, Length: 2, dtype: float64

In [56]: resampled.index
Out[56]: PeriodIndex(['2017Q1', '2017Q3'], dtype='period[2Q-DEC]', freq='2Q-DEC')

Upsampling and calling .ohlc() previously returned a Series, basically identical to calling .asfreq(). OHLC upsampling now returns a DataFrame with columns open, high, low and close (GH13083). This is consistent with downsampling and DatetimeIndex behavior.

Previous Behavior:

In [1]: pi = pd.period_range(start='2000-01-01', freq='D', periods=10)

In [2]: s = pd.Series(np.arange(10), index=pi)

In [3]: s.resample('H').ohlc()
2000-01-01 00:00    0.0
2000-01-10 23:00    NaN
Freq: H, Length: 240, dtype: float64

In [4]: s.resample('M').ohlc()
         open  high  low  close
2000-01     0     9    0      9

New Behavior:

In [57]: pi = pd.period_range(start='2000-01-01', freq='D', periods=10)

In [58]: s = pd.Series(np.arange(10), index=pi)

In [59]: s.resample('H').ohlc()
                  open  high  low  close
2000-01-01 00:00   0.0   0.0  0.0    0.0
2000-01-01 01:00   NaN   NaN  NaN    NaN
2000-01-01 02:00   NaN   NaN  NaN    NaN
2000-01-01 03:00   NaN   NaN  NaN    NaN
2000-01-01 04:00   NaN   NaN  NaN    NaN
2000-01-01 05:00   NaN   NaN  NaN    NaN
2000-01-01 06:00   NaN   NaN  NaN    NaN
...                ...   ...  ...    ...
2000-01-10 17:00   NaN   NaN  NaN    NaN
2000-01-10 18:00   NaN   NaN  NaN    NaN
2000-01-10 19:00   NaN   NaN  NaN    NaN
2000-01-10 20:00   NaN   NaN  NaN    NaN
2000-01-10 21:00   NaN   NaN  NaN    NaN
2000-01-10 22:00   NaN   NaN  NaN    NaN
2000-01-10 23:00   NaN   NaN  NaN    NaN

[240 rows x 4 columns]

In [60]: s.resample('M').ohlc()
         open  high  low  close
2000-01     0     9    0      9

[1 rows x 4 columns]

Improved error handling during item assignment in pd.eval

eval() will now raise a ValueError when item assignment malfunctions, or inplace operations are specified, but there is no item assignment in the expression (GH16732)

In [61]: arr = np.array([1, 2, 3])

Previously, if you attempted the following expression, you would get a not very helpful error message:

In [3]: pd.eval("a = 1 + 2", target=arr, inplace=True)
IndexError: only integers, slices (`:`), ellipsis (`...`), numpy.newaxis (`None`)
and integer or boolean arrays are valid indices

This is a very long way of saying numpy arrays don’t support string-item indexing. With this change, the error message is now this:

In [3]: pd.eval("a = 1 + 2", target=arr, inplace=True)
ValueError: Cannot assign expression output to target

It also used to be possible to evaluate expressions inplace, even if there was no item assignment:

In [4]: pd.eval("1 + 2", target=arr, inplace=True)
Out[4]: 3

However, this input does not make much sense because the output is not being assigned to the target. Now, a ValueError will be raised when such an input is passed in:

In [4]: pd.eval("1 + 2", target=arr, inplace=True)
ValueError: Cannot operate inplace if there is no assignment

Dtype Conversions

Previously assignments, .where() and .fillna() with a bool assignment, would coerce to same the type (e.g. int / float), or raise for datetimelikes. These will now preserve the bools with object dtypes. (GH16821).

In [62]: s = pd.Series([1, 2, 3])
In [5]: s[1] = True

In [6]: s
0    1
1    1
2    3
dtype: int64

New Behavior

In [63]: s[1] = True

In [64]: s
0       1
1    True
2       3
Length: 3, dtype: object

Previously, as assignment to a datetimelike with a non-datetimelike would coerce the non-datetime-like item being assigned (GH14145).

In [65]: s = pd.Series([pd.Timestamp('2011-01-01'), pd.Timestamp('2012-01-01')])
In [1]: s[1] = 1

In [2]: s
0   2011-01-01 00:00:00.000000000
1   1970-01-01 00:00:00.000000001
dtype: datetime64[ns]

These now coerce to object dtype.

In [66]: s[1] = 1

In [67]: s
0    2011-01-01 00:00:00
1                      1
Length: 2, dtype: object
  • Inconsistent behavior in .where() with datetimelikes which would raise rather than coerce to object (GH16402)
  • Bug in assignment against int64 data with np.ndarray with float64 dtype may keep int64 dtype (GH14001)

MultiIndex Constructor with a Single Level

The MultiIndex constructors no longer squeezes a MultiIndex with all length-one levels down to a regular Index. This affects all the MultiIndex constructors. (GH17178)

Previous behavior:

In [2]: pd.MultiIndex.from_tuples([('a',), ('b',)])
Out[2]: Index(['a', 'b'], dtype='object')

Length 1 levels are no longer special-cased. They behave exactly as if you had length 2+ levels, so a MultiIndex is always returned from all of the MultiIndex constructors:

In [68]: pd.MultiIndex.from_tuples([('a',), ('b',)])
MultiIndex(levels=[['a', 'b']],
           codes=[[0, 1]])

UTC Localization with Series

Previously, to_datetime() did not localize datetime Series data when utc=True was passed. Now, to_datetime() will correctly localize Series with a datetime64[ns, UTC] dtype to be consistent with how list-like and Index data are handled. (GH6415).

Previous Behavior

In [69]: s = pd.Series(['20130101 00:00:00'] * 3)
In [12]: pd.to_datetime(s, utc=True)
0   2013-01-01
1   2013-01-01
2   2013-01-01
dtype: datetime64[ns]

New Behavior

In [70]: pd.to_datetime(s, utc=True)
0   2013-01-01 00:00:00+00:00
1   2013-01-01 00:00:00+00:00
2   2013-01-01 00:00:00+00:00
Length: 3, dtype: datetime64[ns, UTC]

Additionally, DataFrames with datetime columns that were parsed by read_sql_table() and read_sql_query() will also be localized to UTC only if the original SQL columns were timezone aware datetime columns.

Consistency of Range Functions

In previous versions, there were some inconsistencies between the various range functions: date_range(), bdate_range(), period_range(), timedelta_range(), and interval_range(). (GH17471).

One of the inconsistent behaviors occurred when the start, end and period parameters were all specified, potentially leading to ambiguous ranges. When all three parameters were passed, interval_range ignored the period parameter, period_range ignored the end parameter, and the other range functions raised. To promote consistency among the range functions, and avoid potentially ambiguous ranges, interval_range and period_range will now raise when all three parameters are passed.

Previous Behavior:

 In [2]: pd.interval_range(start=0, end=4, periods=6)
 IntervalIndex([(0, 1], (1, 2], (2, 3]]

In [3]: pd.period_range(start='2017Q1', end='2017Q4', periods=6, freq='Q')
Out[3]: PeriodIndex(['2017Q1', '2017Q2', '2017Q3', '2017Q4', '2018Q1', '2018Q2'], dtype='period[Q-DEC]', freq='Q-DEC')

New Behavior:

In [2]: pd.interval_range(start=0, end=4, periods=6)
ValueError: Of the three parameters: start, end, and periods, exactly two must be specified

In [3]: pd.period_range(start='2017Q1', end='2017Q4', periods=6, freq='Q')
ValueError: Of the three parameters: start, end, and periods, exactly two must be specified

Additionally, the endpoint parameter end was not included in the intervals produced by interval_range. However, all other range functions include end in their output. To promote consistency among the range functions, interval_range will now include end as the right endpoint of the final interval, except if freq is specified in a way which skips end.

Previous Behavior:

In [4]: pd.interval_range(start=0, end=4)
IntervalIndex([(0, 1], (1, 2], (2, 3]]

New Behavior:

In [71]: pd.interval_range(start=0, end=4)
IntervalIndex([(0, 1], (1, 2], (2, 3], (3, 4]],

No Automatic Matplotlib Converters

Pandas no longer registers our date, time, datetime, datetime64, and Period converters with matplotlib when pandas is imported. Matplotlib plot methods (plt.plot, ax.plot, …), will not nicely format the x-axis for DatetimeIndex or PeriodIndex values. You must explicitly register these methods:

Pandas built-in Series.plot and DataFrame.plot will register these converters on first-use (GH17710).


This change has been temporarily reverted in pandas 0.21.1, for more details see here.

Other API Changes

  • The Categorical constructor no longer accepts a scalar for the categories keyword. (GH16022)
  • Accessing a non-existent attribute on a closed HDFStore will now raise an AttributeError rather than a ClosedFileError (GH16301)
  • read_csv() now issues a UserWarning if the names parameter contains duplicates (GH17095)
  • read_csv() now treats 'null' and 'n/a' strings as missing values by default (GH16471, GH16078)
  • pandas.HDFStore’s string representation is now faster and less detailed. For the previous behavior, use pandas.HDFStore.info(). (GH16503).
  • Compression defaults in HDF stores now follow pytables standards. Default is no compression and if complib is missing and complevel > 0 zlib is used (GH15943)
  • Index.get_indexer_non_unique() now returns a ndarray indexer rather than an Index; this is consistent with Index.get_indexer() (GH16819)
  • Removed the @slow decorator from pandas.util.testing, which caused issues for some downstream packages’ test suites. Use @pytest.mark.slow instead, which achieves the same thing (GH16850)
  • Moved definition of MergeError to the pandas.errors module.
  • The signature of Series.set_axis() and DataFrame.set_axis() has been changed from set_axis(axis, labels) to set_axis(labels, axis=0), for consistency with the rest of the API. The old signature is deprecated and will show a FutureWarning (GH14636)
  • Series.argmin() and Series.argmax() will now raise a TypeError when used with object dtypes, instead of a ValueError (GH13595)
  • Period is now immutable, and will now raise an AttributeError when a user tries to assign a new value to the ordinal or freq attributes (GH17116).
  • to_datetime() when passed a tz-aware origin= kwarg will now raise a more informative ValueError rather than a TypeError (GH16842)
  • to_datetime() now raises a ValueError when format includes %W or %U without also including day of the week and calendar year (GH16774)
  • Renamed non-functional index to index_col in read_stata() to improve API consistency (GH16342)
  • Bug in DataFrame.drop() caused boolean labels False and True to be treated as labels 0 and 1 respectively when dropping indices from a numeric index. This will now raise a ValueError (GH16877)
  • Restricted DateOffset keyword arguments. Previously, DateOffset subclasses allowed arbitrary keyword arguments which could lead to unexpected behavior. Now, only valid arguments will be accepted. (GH17176).


Series.select and DataFrame.select

The Series.select() and DataFrame.select() methods are deprecated in favor of using df.loc[labels.map(crit)] (GH12401)

In [72]: df = pd.DataFrame({'A': [1, 2, 3]}, index=['foo', 'bar', 'baz'])
In [3]: df.select(lambda x: x in ['bar', 'baz'])
FutureWarning: select is deprecated and will be removed in a future release. You can use .loc[crit] as a replacement
bar  2
baz  3
In [73]: df.loc[df.index.map(lambda x: x in ['bar', 'baz'])]
bar  2
baz  3

[2 rows x 1 columns]

Series.argmax and Series.argmin

The behavior of Series.argmax() and Series.argmin() have been deprecated in favor of Series.idxmax() and Series.idxmin(), respectively (GH16830).

For compatibility with NumPy arrays, pd.Series implements argmax and argmin. Since pandas 0.13.0, argmax has been an alias for pandas.Series.idxmax(), and argmin has been an alias for pandas.Series.idxmin(). They return the label of the maximum or minimum, rather than the position.

We’ve deprecated the current behavior of Series.argmax and Series.argmin. Using either of these will emit a FutureWarning. Use Series.idxmax() if you want the label of the maximum. Use Series.values.argmax() if you want the position of the maximum. Likewise for the minimum. In a future release Series.argmax and Series.argmin will return the position of the maximum or minimum.

Removal of prior version deprecations/changes

  • read_excel() has dropped the has_index_names parameter (GH10967)
  • The pd.options.display.height configuration has been dropped (GH3663)
  • The pd.options.display.line_width configuration has been dropped (GH2881)
  • The pd.options.display.mpl_style configuration has been dropped (GH12190)
  • Index has dropped the .sym_diff() method in favor of .symmetric_difference() (GH12591)
  • Categorical has dropped the .order() and .sort() methods in favor of .sort_values() (GH12882)
  • eval() and DataFrame.eval() have changed the default of inplace from None to False (GH11149)
  • The function get_offset_name has been dropped in favor of the .freqstr attribute for an offset (GH11834)
  • pandas no longer tests for compatibility with hdf5-files created with pandas < 0.11 (GH17404).

Performance Improvements

  • Improved performance of instantiating SparseDataFrame (GH16773)
  • Series.dt no longer performs frequency inference, yielding a large speedup when accessing the attribute (GH17210)
  • Improved performance of set_categories() by not materializing the values (GH17508)
  • Timestamp.microsecond no longer re-computes on attribute access (GH17331)
  • Improved performance of the CategoricalIndex for data that is already categorical dtype (GH17513)
  • Improved performance of RangeIndex.min() and RangeIndex.max() by using RangeIndex properties to perform the computations (GH17607)

Documentation Changes

  • Several NaT method docstrings (e.g. NaT.ctime()) were incorrect (GH17327)
  • The documentation has had references to versions < v0.17 removed and cleaned up (GH17442, GH17442, GH17404 & GH17504)

Bug Fixes


  • Bug in assignment against datetime-like data with int may incorrectly convert to datetime-like (GH14145)
  • Bug in assignment against int64 data with np.ndarray with float64 dtype may keep int64 dtype (GH14001)
  • Fixed the return type of IntervalIndex.is_non_overlapping_monotonic to be a Python bool for consistency with similar attributes/methods. Previously returned a numpy.bool_. (GH17237)
  • Bug in IntervalIndex.is_non_overlapping_monotonic when intervals are closed on both sides and overlap at a point (GH16560)
  • Bug in Series.fillna() returns frame when inplace=True and value is dict (GH16156)
  • Bug in Timestamp.weekday_name returning a UTC-based weekday name when localized to a timezone (GH17354)
  • Bug in Timestamp.replace when replacing tzinfo around DST changes (GH15683)
  • Bug in Timedelta construction and arithmetic that would not propagate the Overflow exception (GH17367)
  • Bug in astype() converting to object dtype when passed extension type classes (DatetimeTZDtype, CategoricalDtype) rather than instances. Now a TypeError is raised when a class is passed (GH17780).
  • Bug in to_numeric() in which elements were not always being coerced to numeric when errors='coerce' (GH17007, GH17125)
  • Bug in DataFrame and Series constructors where range objects are converted to int32 dtype on Windows instead of int64 (GH16804)


  • When called with a null slice (e.g. df.iloc[:]), the .iloc and .loc indexers return a shallow copy of the original object. Previously they returned the original object. (GH13873).
  • When called on an unsorted MultiIndex, the loc indexer now will raise UnsortedIndexError only if proper slicing is used on non-sorted levels (GH16734).
  • Fixes regression in 0.20.3 when indexing with a string on a TimedeltaIndex (GH16896).
  • Fixed TimedeltaIndex.get_loc() handling of np.timedelta64 inputs (GH16909).
  • Fix MultiIndex.sort_index() ordering when ascending argument is a list, but not all levels are specified, or are in a different order (GH16934).
  • Fixes bug where indexing with np.inf caused an OverflowError to be raised (GH16957)
  • Bug in reindexing on an empty CategoricalIndex (GH16770)
  • Fixes DataFrame.loc for setting with alignment and tz-aware DatetimeIndex (GH16889)
  • Avoids IndexError when passing an Index or Series to .iloc with older numpy (GH17193)
  • Allow unicode empty strings as placeholders in multilevel columns in Python 2 (GH17099)
  • Bug in .iloc when used with inplace addition or assignment and an int indexer on a MultiIndex causing the wrong indexes to be read from and written to (GH17148)
  • Bug in .isin() in which checking membership in empty Series objects raised an error (GH16991)
  • Bug in CategoricalIndex reindexing in which specified indices containing duplicates were not being respected (GH17323)
  • Bug in intersection of RangeIndex with negative step (GH17296)
  • Bug in IntervalIndex where performing a scalar lookup fails for included right endpoints of non-overlapping monotonic decreasing indexes (GH16417, GH17271)
  • Bug in DataFrame.first_valid_index() and DataFrame.last_valid_index() when no valid entry (GH17400)
  • Bug in Series.rename() when called with a callable, incorrectly alters the name of the Series, rather than the name of the Index. (GH17407)
  • Bug in String.str_get() raises IndexError instead of inserting NaNs when using a negative index. (GH17704)


  • Bug in read_hdf() when reading a timezone aware index from fixed format HDFStore (GH17618)
  • Bug in read_csv() in which columns were not being thoroughly de-duplicated (GH17060)
  • Bug in read_csv() in which specified column names were not being thoroughly de-duplicated (GH17095)
  • Bug in read_csv() in which non integer values for the header argument generated an unhelpful / unrelated error message (GH16338)
  • Bug in read_csv() in which memory management issues in exception handling, under certain conditions, would cause the interpreter to segfault (GH14696, GH16798).
  • Bug in read_csv() when called with low_memory=False in which a CSV with at least one column > 2GB in size would incorrectly raise a MemoryError (GH16798).
  • Bug in read_csv() when called with a single-element list header would return a DataFrame of all NaN values (GH7757)
  • Bug in DataFrame.to_csv() defaulting to ‘ascii’ encoding in Python 3, instead of ‘utf-8’ (GH17097)
  • Bug in read_stata() where value labels could not be read when using an iterator (GH16923)
  • Bug in read_stata() where the index was not set (GH16342)
  • Bug in read_html() where import check fails when run in multiple threads (GH16928)
  • Bug in read_csv() where automatic delimiter detection caused a TypeError to be thrown when a bad line was encountered rather than the correct error message (GH13374)
  • Bug in DataFrame.to_html() with notebook=True where DataFrames with named indices or non-MultiIndex indices had undesired horizontal or vertical alignment for column or row labels, respectively (GH16792)
  • Bug in DataFrame.to_html() in which there was no validation of the justify parameter (GH17527)
  • Bug in HDFStore.select() when reading a contiguous mixed-data table featuring VLArray (GH17021)
  • Bug in to_json() where several conditions (including objects with unprintable symbols, objects with deep recursion, overlong labels) caused segfaults instead of raising the appropriate exception (GH14256)


  • Bug in plotting methods using secondary_y and fontsize not setting secondary axis font size (GH12565)
  • Bug when plotting timedelta and datetime dtypes on y-axis (GH16953)
  • Line plots no longer assume monotonic x data when calculating xlims, they show the entire lines now even for unsorted x data. (GH11310, GH11471)
  • With matplotlib 2.0.0 and above, calculation of x limits for line plots is left to matplotlib, so that its new default settings are applied. (GH15495)
  • Bug in Series.plot.bar or DataFrame.plot.bar with y not respecting user-passed color (GH16822)
  • Bug causing plotting.parallel_coordinates to reset the random seed when using random colors (GH17525)


  • Bug in DataFrame.resample(...).size() where an empty DataFrame did not return a Series (GH14962)
  • Bug in infer_freq() causing indices with 2-day gaps during the working week to be wrongly inferred as business daily (GH16624)
  • Bug in .rolling(...).quantile() which incorrectly used different defaults than Series.quantile() and DataFrame.quantile() (GH9413, GH16211)
  • Bug in groupby.transform() that would coerce boolean dtypes back to float (GH16875)
  • Bug in Series.resample(...).apply() where an empty Series modified the source index and did not return the name of a Series (GH14313)
  • Bug in .rolling(...).apply(...) with a DataFrame with a DatetimeIndex, a window of a timedelta-convertible and min_periods >= 1 (GH15305)
  • Bug in DataFrame.groupby where index and column keys were not recognized correctly when the number of keys equaled the number of elements on the groupby axis (GH16859)
  • Bug in groupby.nunique() with TimeGrouper which cannot handle NaT correctly (GH17575)
  • Bug in DataFrame.groupby where a single level selection from a MultiIndex unexpectedly sorts (GH17537)
  • Bug in DataFrame.groupby where spurious warning is raised when Grouper object is used to override ambiguous column name (GH17383)
  • Bug in TimeGrouper differs when passes as a list and as a scalar (GH17530)


  • Bug in SparseSeries raises AttributeError when a dictionary is passed in as data (GH16905)
  • Bug in SparseDataFrame.fillna() not filling all NaNs when frame was instantiated from SciPy sparse matrix (GH16112)
  • Bug in SparseSeries.unstack() and SparseDataFrame.stack() (GH16614, GH15045)
  • Bug in make_sparse() treating two numeric/boolean data, which have same bits, as same when array dtype is object (GH17574)
  • SparseArray.all() and SparseArray.any() are now implemented to handle SparseArray, these were used but not implemented (GH17570)


  • Joining/Merging with a non unique PeriodIndex raised a TypeError (GH16871)
  • Bug in crosstab() where non-aligned series of integers were casted to float (GH17005)
  • Bug in merging with categorical dtypes with datetimelikes incorrectly raised a TypeError (GH16900)
  • Bug when using isin() on a large object series and large comparison array (GH16012)
  • Fixes regression from 0.20, Series.aggregate() and DataFrame.aggregate() allow dictionaries as return values again (GH16741)
  • Fixes dtype of result with integer dtype input, from pivot_table() when called with margins=True (GH17013)
  • Bug in crosstab() where passing two Series with the same name raised a KeyError (GH13279)
  • Series.argmin(), Series.argmax(), and their counterparts on DataFrame and groupby objects work correctly with floating point data that contains infinite values (GH13595).
  • Bug in unique() where checking a tuple of strings raised a TypeError (GH17108)
  • Bug in concat() where order of result index was unpredictable if it contained non-comparable elements (GH17344)
  • Fixes regression when sorting by multiple columns on a datetime64 dtype Series with NaT values (GH16836)
  • Bug in pivot_table() where the result’s columns did not preserve the categorical dtype of columns when dropna was False (GH17842)
  • Bug in DataFrame.drop_duplicates where dropping with non-unique column names raised a ValueError (GH17836)
  • Bug in unstack() which, when called on a list of levels, would discard the fillna argument (GH13971)
  • Bug in the alignment of range objects and other list-likes with DataFrame leading to operations being performed row-wise instead of column-wise (GH17901)


  • Bug in .clip() with axis=1 and a list-like for threshold is passed; previously this raised ValueError (GH15390)
  • Series.clip() and DataFrame.clip() now treat NA values for upper and lower arguments as None instead of raising ValueError (GH17276).


  • Bug in Series.isin() when called with a categorical (GH16639)
  • Bug in the categorical constructor with empty values and categories causing the .categories to be an empty Float64Index rather than an empty Index with object dtype (GH17248)
  • Bug in categorical operations with Series.cat not preserving the original Series’ name (GH17509)
  • Bug in DataFrame.merge() failing for categorical columns with boolean/int data types (GH17187)
  • Bug in constructing a Categorical/CategoricalDtype when the specified categories are of categorical type (GH17884).


  • Compatibility with PyPy in read_csv() with usecols=[<unsorted ints>] and read_json() (GH17351)
  • Split tests into cases for CPython and PyPy where needed, which highlights the fragility of index matching with float('nan'), np.nan and NAT (GH17351)
  • Fix DataFrame.memory_usage() to support PyPy. Objects on PyPy do not have a fixed size, so an approximation is used instead (GH17228)


  • Bug where some inplace operators were not being wrapped and produced a copy when invoked (GH12962)
  • Bug in eval() where the inplace parameter was being incorrectly handled (GH16732)


A total of 206 people contributed patches to this release. People with a “+” by their names contributed a patch for the first time.

  • 3553x +
  • Aaron Barber
  • Adam Gleave +
  • Adam Smith +
  • AdamShamlian +
  • Adrian Liaw +
  • Alan Velasco +
  • Alan Yee +
  • Alex B +
  • Alex Lubbock +
  • Alex Marchenko +
  • Alex Rychyk +
  • Amol K +
  • Andreas Winkler
  • Andrew +
  • Andrew 亮
  • André Jonasson +
  • Becky Sweger
  • Berkay +
  • Bob Haffner +
  • Bran Yang
  • Brian Tu +
  • Brock Mendel +
  • Carol Willing +
  • Carter Green +
  • Chankey Pathak +
  • Chris
  • Chris Billington
  • Chris Filo Gorgolewski +
  • Chris Kerr
  • Chris M +
  • Chris Mazzullo +
  • Christian Prinoth
  • Christian Stade-Schuldt
  • Christoph Moehl +
  • DSM
  • Daniel Chen +
  • Daniel Grady
  • Daniel Himmelstein
  • Dave Willmer
  • David Cook
  • David Gwynne
  • David Read +
  • Dillon Niederhut +
  • Douglas Rudd
  • Eric Stein +
  • Eric Wieser +
  • Erik Fredriksen
  • Florian Wilhelm +
  • Floris Kint +
  • Forbidden Donut
  • Gabe F +
  • Giftlin +
  • Giftlin Rajaiah +
  • Giulio Pepe +
  • Guilherme Beltramini
  • Guillem Borrell +
  • Hanmin Qin +
  • Hendrik Makait +
  • Hugues Valois
  • Hussain Tamboli +
  • Iva Miholic +
  • Jan Novotný +
  • Jan Rudolph
  • Jean Helie +
  • Jean-Baptiste Schiratti +
  • Jean-Mathieu Deschenes
  • Jeff Knupp +
  • Jeff Reback
  • Jeff Tratner
  • JennaVergeynst
  • JimStearns206
  • Joel Nothman
  • John W. O’Brien
  • Jon Crall +
  • Jon Mease
  • Jonathan J. Helmus +
  • Joris Van den Bossche
  • JosephWagner
  • Juarez Bochi
  • Julian Kuhlmann +
  • Karel De Brabandere
  • Kassandra Keeton +
  • Keiron Pizzey +
  • Keith Webber
  • Kernc
  • Kevin Sheppard
  • Kirk Hansen +
  • Licht Takeuchi +
  • Lucas Kushner +
  • Mahdi Ben Jelloul +
  • Makarov Andrey +
  • Malgorzata Turzanska +
  • Marc Garcia +
  • Margaret Sy +
  • MarsGuy +
  • Matt Bark +
  • Matthew Roeschke
  • Matti Picus
  • Mehmet Ali “Mali” Akmanalp
  • Michael Gasvoda +
  • Michael Penkov +
  • Milo +
  • Morgan Stuart +
  • Morgan243 +
  • Nathan Ford +
  • Nick Eubank
  • Nick Garvey +
  • Oleg Shteynbuk +
  • P-Tillmann +
  • Pankaj Pandey
  • Patrick Luo
  • Patrick O’Melveny
  • Paul Reidy +
  • Paula +
  • Peter Quackenbush
  • Peter Yanovich +
  • Phillip Cloud
  • Pierre Haessig
  • Pietro Battiston
  • Pradyumna Reddy Chinthala
  • Prasanjit Prakash
  • RobinFiveWords
  • Ryan Hendrickson
  • Sam Foo
  • Sangwoong Yoon +
  • Simon Gibbons +
  • SimonBaron
  • Steven Cutting +
  • Sudeep +
  • Sylvia +
  • T N +
  • Telt
  • Thomas A Caswell
  • Tim Swast +
  • Tom Augspurger
  • Tong SHEN
  • Tuan +
  • Utkarsh Upadhyay +
  • Vincent La +
  • Vivek +
  • WANG Aiyong
  • WBare
  • Wes McKinney
  • XF +
  • Yi Liu +
  • Yosuke Nakabayashi +
  • aaron315 +
  • abarber4gh +
  • aernlund +
  • agustín méndez +
  • andymaheshw +
  • ante328 +
  • aviolov +
  • bpraggastis
  • cbertinato +
  • cclauss +
  • chernrick
  • chris-b1
  • dkamm +
  • dwkenefick
  • economy
  • faic +
  • fding253 +
  • gfyoung
  • guygoldberg +
  • hhuuggoo +
  • huashuai +
  • ian
  • iulia +
  • jaredsnyder
  • jbrockmendel +
  • jdeschenes
  • jebob +
  • jschendel +
  • keitakurita
  • kernc +
  • kiwirob +
  • kjford
  • linebp
  • lloydkirk
  • louispotok +
  • majiang +
  • manikbhandari +
  • matthiashuschle +
  • mattip
  • maxwasserman +
  • mjlove12 +
  • nmartensen +
  • pandas-docs-bot +
  • parchd-1 +
  • philipphanemann +
  • rdk1024 +
  • reidy-p +
  • ri938
  • ruiann +
  • rvernica +
  • s-weigand +
  • scotthavard92 +
  • skwbc +
  • step4me +
  • tobycheese +
  • topper-123 +
  • tsdlovell
  • ysau +
  • zzgao +
Scroll To Top