Table Of Contents


Enter search terms or a module, class or function name.

What’s new in 0.25.0 (July 18, 2019)


Starting with the 0.25.x series of releases, pandas only supports Python 3.5.3 and higher. See Dropping Python 2.7 for more details.


The minimum supported Python version will be bumped to 3.6 in a future release.


Panel has been fully removed. For N-D labeled data structures, please use xarray


read_pickle() and read_msgpack() are only guaranteed backwards compatible back to pandas version 0.20.3 (GH27082)

These are the changes in pandas 0.25.0. See Release Notes for a full changelog including other versions of pandas.


Groupby aggregation with relabeling

Pandas has added special groupby behavior, known as “named aggregation”, for naming the output columns when applying multiple aggregation functions to specific columns (GH18366, GH26512).

In [1]: animals = pd.DataFrame({'kind': ['cat', 'dog', 'cat', 'dog'],
   ...:                         'height': [9.1, 6.0, 9.5, 34.0],
   ...:                         'weight': [7.9, 7.5, 9.9, 198.0]})

In [2]: animals
  kind  height  weight
0  cat     9.1     7.9
1  dog     6.0     7.5
2  cat     9.5     9.9
3  dog    34.0   198.0

[4 rows x 3 columns]

In [3]: animals.groupby("kind").agg(
   ...:     min_height=pd.NamedAgg(column='height', aggfunc='min'),
   ...:     max_height=pd.NamedAgg(column='height', aggfunc='max'),
   ...:     average_weight=pd.NamedAgg(column='weight', aggfunc=np.mean),
   ...: )
      min_height  max_height  average_weight
cat          9.1         9.5            8.90
dog          6.0        34.0          102.75

[2 rows x 3 columns]

Pass the desired columns names as the **kwargs to .agg. The values of **kwargs should be tuples where the first element is the column selection, and the second element is the aggregation function to apply. Pandas provides the pandas.NamedAgg namedtuple to make it clearer what the arguments to the function are, but plain tuples are accepted as well.

In [4]: animals.groupby("kind").agg(
   ...:     min_height=('height', 'min'),
   ...:     max_height=('height', 'max'),
   ...:     average_weight=('weight', np.mean),
   ...: )
      min_height  max_height  average_weight
cat          9.1         9.5            8.90
dog          6.0        34.0          102.75

[2 rows x 3 columns]

Named aggregation is the recommended replacement for the deprecated “dict-of-dicts” approach to naming the output of column-specific aggregations (Deprecate groupby.agg() with a dictionary when renaming).

A similar approach is now available for Series groupby objects as well. Because there’s no need for column selection, the values can just be the functions to apply

In [5]: animals.groupby("kind").height.agg(
   ...:     min_height="min",
   ...:     max_height="max",
   ...: )
      min_height  max_height
cat          9.1         9.5
dog          6.0        34.0

[2 rows x 2 columns]

This type of aggregation is the recommended alternative to the deprecated behavior when passing a dict to a Series groupby aggregation (Deprecate groupby.agg() with a dictionary when renaming).

See Named aggregation for more.

Groupby Aggregation with multiple lambdas

You can now provide multiple lambda functions to a list-like aggregation in pandas.core.groupby.GroupBy.agg (GH26430).

In [6]: animals.groupby('kind').height.agg([
   ...:     lambda x: x.iloc[0], lambda x: x.iloc[-1]
   ...: ])
      <lambda_0>  <lambda_1>
cat          9.1         9.5
dog          6.0        34.0

[2 rows x 2 columns]

In [7]: animals.groupby('kind').agg([
   ...:     lambda x: x.iloc[0] - x.iloc[1],
   ...:     lambda x: x.iloc[0] + x.iloc[1]
   ...: ])
         height                weight           
     <lambda_0> <lambda_1> <lambda_0> <lambda_1>
cat        -0.4       18.6       -2.0       17.8
dog       -28.0       40.0     -190.5      205.5

[2 rows x 4 columns]

Previously, these raised a SpecificationError.

Better repr for MultiIndex

Printing of MultiIndex instances now shows tuples of each row and ensures that the tuple items are vertically aligned, so it’s now easier to understand the structure of the MultiIndex. (GH13480):

The repr now looks like this:

In [8]: pd.MultiIndex.from_product([['a', 'abc'], range(500)])
MultiIndex([(  'a',   0),
            (  'a',   1),
            (  'a',   2),
            (  'a',   3),
            (  'a',   4),
            (  'a',   5),
            (  'a',   6),
            (  'a',   7),
            (  'a',   8),
            (  'a',   9),
            ('abc', 490),
            ('abc', 491),
            ('abc', 492),
            ('abc', 493),
            ('abc', 494),
            ('abc', 495),
            ('abc', 496),
            ('abc', 497),
            ('abc', 498),
            ('abc', 499)],

Previously, outputting a MultiIndex printed all the levels and codes of the MultiIndex, which was visually unappealing and made the output more difficult to navigate. For example (limiting the range to 5):

In [1]: pd.MultiIndex.from_product([['a', 'abc'], range(5)])
Out[1]: MultiIndex(levels=[['a', 'abc'], [0, 1, 2, 3]],
   ...:            codes=[[0, 0, 0, 0, 1, 1, 1, 1], [0, 1, 2, 3, 0, 1, 2, 3]])

In the new repr, all values will be shown, if the number of rows is smaller than options.display.max_seq_items (default: 100 items). Horizontally, the output will truncate, if it’s wider than options.display.width (default: 80 characters).

Shorter truncated repr for Series and DataFrame

Currently, the default display options of pandas ensure that when a Series or DataFrame has more than 60 rows, its repr gets truncated to this maximum of 60 rows (the display.max_rows option). However, this still gives a repr that takes up a large part of the vertical screen estate. Therefore, a new option display.min_rows is introduced with a default of 10 which determines the number of rows showed in the truncated repr:

  • For small Series or DataFrames, up to max_rows number of rows is shown (default: 60).
  • For larger Series of DataFrame with a length above max_rows, only min_rows number of rows is shown (default: 10, i.e. the first and last 5 rows).

This dual option allows to still see the full content of relatively small objects (e.g. df.head(20) shows all 20 rows), while giving a brief repr for large objects.

To restore the previous behaviour of a single threshold, set pd.options.display.min_rows = None.

Json normalize with max_level param support

json_normalize() normalizes the provided input dict to all nested levels. The new max_level parameter provides more control over which level to end normalization (GH23843):

The repr now looks like this:

In [9]: from import json_normalize

In [10]: data = [{
   ....:     'CreatedBy': {'Name': 'User001'},
   ....:     'Lookup': {'TextField': 'Some text',
   ....:                'UserField': {'Id': 'ID001', 'Name': 'Name001'}},
   ....:     'Image': {'a': 'b'}
   ....: }]

In [11]: json_normalize(data, max_level=1)
  CreatedBy.Name Lookup.TextField                    Lookup.UserField Image.a
0        User001        Some text  {'Id': 'ID001', 'Name': 'Name001'}       b

[1 rows x 4 columns]

Series.explode to split list-like values to rows

Series and DataFrame have gained the DataFrame.explode() methods to transform list-likes to individual rows. See section on Exploding list-like column in docs for more information (GH16538, GH10511)

Here is a typical usecase. You have comma separated string in a column.

In [12]: df = pd.DataFrame([{'var1': 'a,b,c', 'var2': 1},
   ....:                    {'var1': 'd,e,f', 'var2': 2}])

In [13]: df
    var1  var2
0  a,b,c     1
1  d,e,f     2

[2 rows x 2 columns]

Creating a long form DataFrame is now straightforward using chained operations

In [14]: df.assign(var1=df.var1.str.split(',')).explode('var1')
  var1  var2
0    a     1
0    b     1
0    c     1
1    d     2
1    e     2
1    f     2

[6 rows x 2 columns]

Other enhancements

  • DataFrame.plot() keywords logy, logx and loglog can now accept the value 'sym' for symlog scaling. (GH24867)
  • Added support for ISO week year format (‘%G-%V-%u’) when parsing datetimes using to_datetime() (GH16607)
  • Indexing of DataFrame and Series now accepts zerodim np.ndarray (GH24919)
  • Timestamp.replace() now supports the fold argument to disambiguate DST transition times (GH25017)
  • DataFrame.at_time() and Series.at_time() now support datetime.time objects with timezones (GH24043)
  • DataFrame.pivot_table() now accepts an observed parameter which is passed to underlying calls to DataFrame.groupby() to speed up grouping categorical data. (GH24923)
  • Series.str has gained Series.str.casefold() method to removes all case distinctions present in a string (GH25405)
  • DataFrame.set_index() now works for instances of abc.Iterator, provided their output is of the same length as the calling frame (GH22484, GH24984)
  • DatetimeIndex.union() now supports the sort argument. The behavior of the sort parameter matches that of Index.union() (GH24994)
  • RangeIndex.union() now supports the sort argument. If sort=False an unsorted Int64Index is always returned. sort=None is the default and returns a monotonically increasing RangeIndex if possible or a sorted Int64Index if not (GH24471)
  • TimedeltaIndex.intersection() now also supports the sort keyword (GH24471)
  • DataFrame.rename() now supports the errors argument to raise errors when attempting to rename nonexistent keys (GH13473)
  • Added Sparse accessor for working with a DataFrame whose values are sparse (GH25681)
  • RangeIndex has gained start, stop, and step attributes (GH25710)
  • datetime.timezone objects are now supported as arguments to timezone methods and constructors (GH25065)
  • DataFrame.query() and DataFrame.eval() now supports quoting column names with backticks to refer to names with spaces (GH6508)
  • merge_asof() now gives a more clear error message when merge keys are categoricals that are not equal (GH26136)
  • pandas.core.window.Rolling() supports exponential (or Poisson) window type (GH21303)
  • Error message for missing required imports now includes the original import error’s text (GH23868)
  • DatetimeIndex and TimedeltaIndex now have a mean method (GH24757)
  • DataFrame.describe() now formats integer percentiles without decimal point (GH26660)
  • Added support for reading SPSS .sav files using read_spss() (GH26537)
  • Added new option plotting.backend to be able to select a plotting backend different than the existing matplotlib one. Use pandas.set_option('plotting.backend', '<backend-module>') where <backend-module is a library implementing the pandas plotting API (GH14130)
  • pandas.offsets.BusinessHour supports multiple opening hours intervals (GH15481)
  • read_excel() can now use openpyxl to read Excel files via the engine='openpyxl' argument. This will become the default in a future release (GH11499)
  • supports reading OpenDocument tables. Specify engine='odf' to enable. Consult the IO User Guide for more details (GH9070)
  • Interval, IntervalIndex, and IntervalArray have gained an is_empty attribute denoting if the given interval(s) are empty (GH27219)

Backwards incompatible API changes

Indexing with date strings with UTC offsets

Indexing a DataFrame or Series with a DatetimeIndex with a date string with a UTC offset would previously ignore the UTC offset. Now, the UTC offset is respected in indexing. (GH24076, GH16785)

In [15]: df = pd.DataFrame([0], index=pd.DatetimeIndex(['2019-01-01'], tz='US/Pacific'))

In [16]: df
2019-01-01 00:00:00-08:00  0

[1 rows x 1 columns]

Previous behavior:

In [3]: df['2019-01-01 00:00:00+04:00':'2019-01-01 01:00:00+04:00']
2019-01-01 00:00:00-08:00  0

New behavior:

In [17]: df['2019-01-01 12:00:00+04:00':'2019-01-01 13:00:00+04:00']
2019-01-01 00:00:00-08:00  0

[1 rows x 1 columns]

MultiIndex constructed from levels and codes

Constructing a MultiIndex with NaN levels or codes value < -1 was allowed previously. Now, construction with codes value < -1 is not allowed and NaN levels’ corresponding codes would be reassigned as -1. (GH19387)

Previous behavior:

In [1]: pd.MultiIndex(levels=[[np.nan, None, pd.NaT, 128, 2]],
   ...:               codes=[[0, -1, 1, 2, 3, 4]])
Out[1]: MultiIndex(levels=[[nan, None, NaT, 128, 2]],
                   codes=[[0, -1, 1, 2, 3, 4]])

In [2]: pd.MultiIndex(levels=[[1, 2]], codes=[[0, -2]])
Out[2]: MultiIndex(levels=[[1, 2]],
                   codes=[[0, -2]])

New behavior:

In [18]: pd.MultiIndex(levels=[[np.nan, None, pd.NaT, 128, 2]],
   ....:               codes=[[0, -1, 1, 2, 3, 4]])
            (  2,)],

In [19]: pd.MultiIndex(levels=[[1, 2]], codes=[[0, -2]])
ValueError                                Traceback (most recent call last)
<ipython-input-19-225a01af3975> in <module>
----> 1 pd.MultiIndex(levels=[[1, 2]], codes=[[0, -2]])

/pandas/pandas/util/ in wrapper(*args, **kwargs)
    206                 else:
    207                     kwargs[new_arg_name] = new_arg_value
--> 208             return func(*args, **kwargs)
    210         return wrapper

/pandas/pandas/core/indexes/ in __new__(cls, levels, codes, sortorder, names, dtype, copy, name, verify_integrity, _set_identity)
    271         if verify_integrity:
--> 272             new_codes = result._verify_integrity()
    273             result._codes = new_codes

/pandas/pandas/core/indexes/ in _verify_integrity(self, codes, levels)
    348                 raise ValueError(
    349                     "On level {level}, code value ({code})"
--> 350                     " < -1".format(level=i, code=level_codes.min())
    351                 )
    352             if not level.is_unique:

ValueError: On level 0, code value (-2) < -1

Groupby.apply on DataFrame evaluates first group only once

The implementation of DataFrameGroupBy.apply() previously evaluated the supplied function consistently twice on the first group to infer if it is safe to use a fast code path. Particularly for functions with side effects, this was an undesired behavior and may have led to surprises. (GH2936, GH2656, GH7739, GH10519, GH12155, GH20084, GH21417)

Now every group is evaluated only a single time.

In [20]: df = pd.DataFrame({"a": ["x", "y"], "b": [1, 2]})

In [21]: df
   a  b
0  x  1
1  y  2

[2 rows x 2 columns]

In [22]: def func(group):
   ....:     print(
   ....:     return group

Previous behavior:

In [3]: df.groupby('a').apply(func)
   a  b
0  x  1
1  y  2

New behavior:

In [23]: df.groupby("a").apply(func)
   a  b
0  x  1
1  y  2

[2 rows x 2 columns]

Concatenating sparse values

When passed DataFrames whose values are sparse, concat() will now return a Series or DataFrame with sparse values, rather than a SparseDataFrame (GH25702).

In [24]: df = pd.DataFrame({"A": pd.SparseArray([0, 1])})

Previous behavior:

In [2]: type(pd.concat([df, df]))

New behavior:

In [25]: type(pd.concat([df, df]))
Out[25]: pandas.core.frame.DataFrame

This now matches the existing behavior of concat on Series with sparse values. concat() will continue to return a SparseDataFrame when all the values are instances of SparseDataFrame.

This change also affects routines using concat() internally, like get_dummies(), which now returns a DataFrame in all cases (previously a SparseDataFrame was returned if all the columns were dummy encoded, and a DataFrame otherwise).

Providing any SparseSeries or SparseDataFrame to concat() will cause a SparseSeries or SparseDataFrame to be returned, as before.

The .str-accessor performs stricter type checks

Due to the lack of more fine-grained dtypes, Series.str so far only checked whether the data was of object dtype. Series.str will now infer the dtype data within the Series; in particular, 'bytes'-only data will raise an exception (except for Series.str.decode(), Series.str.get(), Series.str.len(), Series.str.slice()), see GH23163, GH23011, GH23551.

Previous behavior:

In [1]: s = pd.Series(np.array(['a', 'ba', 'cba'], 'S'), dtype=object)

In [2]: s
0      b'a'
1     b'ba'
2    b'cba'
dtype: object

In [3]: s.str.startswith(b'a')
0     True
1    False
2    False
dtype: bool

New behavior:

In [26]: s = pd.Series(np.array(['a', 'ba', 'cba'], 'S'), dtype=object)

In [27]: s
0      b'a'
1     b'ba'
2    b'cba'
Length: 3, dtype: object

In [28]: s.str.startswith(b'a')
TypeError                                 Traceback (most recent call last)
<ipython-input-28-ac784692b361> in <module>
----> 1 s.str.startswith(b'a')

/pandas/pandas/core/ in wrapper(self, *args, **kwargs)
   1840                     )
   1841                 )
-> 1842                 raise TypeError(msg)
   1843             return func(self, *args, **kwargs)

TypeError: Cannot use .str.startswith with values of inferred dtype 'bytes'.

Categorical dtypes are preserved during groupby

Previously, columns that were categorical, but not the groupby key(s) would be converted to object dtype during groupby operations. Pandas now will preserve these dtypes. (GH18502)

In [29]: cat = pd.Categorical(["foo", "bar", "bar", "qux"], ordered=True)

In [30]: df = pd.DataFrame({'payload': [-1, -2, -1, -2], 'col': cat})

In [31]: df
   payload  col
0       -1  foo
1       -2  bar
2       -1  bar
3       -2  qux

[4 rows x 2 columns]

In [32]: df.dtypes
payload       int64
col        category
Length: 2, dtype: object

Previous Behavior:

In [5]: df.groupby('payload').first().col.dtype
Out[5]: dtype('O')

New Behavior:

In [33]: df.groupby('payload').first().col.dtype
Out[33]: CategoricalDtype(categories=['bar', 'foo', 'qux'], ordered=True)

Incompatible Index type unions

When performing Index.union() operations between objects of incompatible dtypes, the result will be a base Index of dtype object. This behavior holds true for unions between Index objects that previously would have been prohibited. The dtype of empty Index objects will now be evaluated before performing union operations rather than simply returning the other Index object. Index.union() can now be considered commutative, such that A.union(B) == B.union(A) (GH23525).

Previous behavior:

In [1]: pd.period_range('19910905', periods=2).union(pd.Int64Index([1, 2, 3]))
ValueError: can only call with other PeriodIndex-ed objects

In [2]: pd.Index([], dtype=object).union(pd.Index([1, 2, 3]))
Out[2]: Int64Index([1, 2, 3], dtype='int64')

New behavior:

In [34]: pd.period_range('19910905', periods=2).union(pd.Int64Index([1, 2, 3]))
Out[34]: Index([1991-09-05, 1991-09-06, 1, 2, 3], dtype='object')

In [35]: pd.Index([], dtype=object).union(pd.Index([1, 2, 3]))
Out[35]: Index([1, 2, 3], dtype='object')

Note that integer- and floating-dtype indexes are considered “compatible”. The integer values are coerced to floating point, which may result in loss of precision. See Set operations on Index objects for more.

DataFrame groupby ffill/bfill no longer return group labels

The methods ffill, bfill, pad and backfill of DataFrameGroupBy previously included the group labels in the return value, which was inconsistent with other groupby transforms. Now only the filled values are returned. (GH21521)

In [36]: df = pd.DataFrame({"a": ["x", "y"], "b": [1, 2]})

In [37]: df
   a  b
0  x  1
1  y  2

[2 rows x 2 columns]

Previous behavior:

In [3]: df.groupby("a").ffill()
   a  b
0  x  1
1  y  2

New behavior:

In [38]: df.groupby("a").ffill()
0  1
1  2

[2 rows x 1 columns]

DataFrame describe on an empty categorical / object column will return top and freq

When calling DataFrame.describe() with an empty categorical / object column, the ‘top’ and ‘freq’ columns were previously omitted, which was inconsistent with the output for non-empty columns. Now the ‘top’ and ‘freq’ columns will always be included, with numpy.nan in the case of an empty DataFrame (GH26397)

In [39]: df = pd.DataFrame({"empty_col": pd.Categorical([])})

In [40]: df
Empty DataFrame
Columns: [empty_col]
Index: []

[0 rows x 1 columns]

Previous behavior:

In [3]: df.describe()
count           0
unique          0

New behavior:

In [41]: df.describe()
count          0
unique         0
top          NaN
freq         NaN

[4 rows x 1 columns]

__str__ methods now call __repr__ rather than vice versa

Pandas has until now mostly defined string representations in a Pandas objects’s __str__/__unicode__/__bytes__ methods, and called __str__ from the __repr__ method, if a specific __repr__ method is not found. This is not needed for Python3. In Pandas 0.25, the string representations of Pandas objects are now generally defined in __repr__, and calls to __str__ in general now pass the call on to the __repr__, if a specific __str__ method doesn’t exist, as is standard for Python. This change is backward compatible for direct usage of Pandas, but if you subclass Pandas objects and give your subclasses specific __str__/__repr__ methods, you may have to adjust your __str__/__repr__ methods (GH26495).

Indexing an IntervalIndex with Interval objects

Indexing methods for IntervalIndex have been modified to require exact matches only for Interval queries. IntervalIndex methods previously matched on any overlapping Interval. Behavior with scalar points, e.g. querying with an integer, is unchanged (GH16316).

In [42]: ii = pd.IntervalIndex.from_tuples([(0, 4), (1, 5), (5, 8)])

In [43]: ii
IntervalIndex([(0, 4], (1, 5], (5, 8]],

The in operator (__contains__) now only returns True for exact matches to Intervals in the IntervalIndex, whereas this would previously return True for any Interval overlapping an Interval in the IntervalIndex.

Previous behavior:

In [4]: pd.Interval(1, 2, closed='neither') in ii
Out[4]: True

In [5]: pd.Interval(-10, 10, closed='both') in ii
Out[5]: True

New behavior:

In [44]: pd.Interval(1, 2, closed='neither') in ii
Out[44]: False

In [45]: pd.Interval(-10, 10, closed='both') in ii
Out[45]: False

The get_loc() method now only returns locations for exact matches to Interval queries, as opposed to the previous behavior of returning locations for overlapping matches. A KeyError will be raised if an exact match is not found.

Previous behavior:

In [6]: ii.get_loc(pd.Interval(1, 5))
Out[6]: array([0, 1])

In [7]: ii.get_loc(pd.Interval(2, 6))
Out[7]: array([0, 1, 2])

New behavior:

In [6]: ii.get_loc(pd.Interval(1, 5))
Out[6]: 1

In [7]: ii.get_loc(pd.Interval(2, 6))
KeyError: Interval(2, 6, closed='right')

Likewise, get_indexer() and get_indexer_non_unique() will also only return locations for exact matches to Interval queries, with -1 denoting that an exact match was not found.

These indexing changes extend to querying a Series or DataFrame with an IntervalIndex index.

In [46]: s = pd.Series(list('abc'), index=ii)

In [47]: s
(0, 4]    a
(1, 5]    b
(5, 8]    c
Length: 3, dtype: object

Selecting from a Series or DataFrame using [] (__getitem__) or loc now only returns exact matches for Interval queries.

Previous behavior:

In [8]: s[pd.Interval(1, 5)]
(0, 4]    a
(1, 5]    b
dtype: object

In [9]: s.loc[pd.Interval(1, 5)]
(0, 4]    a
(1, 5]    b
dtype: object

New behavior:

In [48]: s[pd.Interval(1, 5)]
Out[48]: 'b'

In [49]: s.loc[pd.Interval(1, 5)]
Out[49]: 'b'

Similarly, a KeyError will be raised for non-exact matches instead of returning overlapping matches.

Previous behavior:

In [9]: s[pd.Interval(2, 3)]
(0, 4]    a
(1, 5]    b
dtype: object

In [10]: s.loc[pd.Interval(2, 3)]
(0, 4]    a
(1, 5]    b
dtype: object

New behavior:

In [6]: s[pd.Interval(2, 3)]
KeyError: Interval(2, 3, closed='right')

In [7]: s.loc[pd.Interval(2, 3)]
KeyError: Interval(2, 3, closed='right')

The overlaps() method can be used to create a boolean indexer that replicates the previous behavior of returning overlapping matches.

New behavior:

In [50]: idxr = s.index.overlaps(pd.Interval(2, 3))

In [51]: idxr
Out[51]: array([ True,  True, False])

In [52]: s[idxr]
(0, 4]    a
(1, 5]    b
Length: 2, dtype: object

In [53]: s.loc[idxr]
(0, 4]    a
(1, 5]    b
Length: 2, dtype: object

Binary ufuncs on Series now align

Applying a binary ufunc like numpy.power() now aligns the inputs when both are Series (GH23293).

In [54]: s1 = pd.Series([1, 2, 3], index=['a', 'b', 'c'])

In [55]: s2 = pd.Series([3, 4, 5], index=['d', 'c', 'b'])

In [56]: s1
a    1
b    2
c    3
Length: 3, dtype: int64

In [57]: s2
d    3
c    4
b    5
Length: 3, dtype: int64

Previous behavior

In [5]: np.power(s1, s2)
a      1
b     16
c    243
dtype: int64

New behavior

In [58]: np.power(s1, s2)
a     1.0
b    32.0
c    81.0
d     NaN
Length: 4, dtype: float64

This matches the behavior of other binary operations in pandas, like Series.add(). To retain the previous behavior, convert the other Series to an array before applying the ufunc.

In [59]: np.power(s1, s2.array)
a      1
b     16
c    243
Length: 3, dtype: int64

Categorical.argsort now places missing values at the end

Categorical.argsort() now places missing values at the end of the array, making it consistent with NumPy and the rest of pandas (GH21801).

In [60]: cat = pd.Categorical(['b', None, 'a'], categories=['a', 'b'], ordered=True)

Previous behavior

In [2]: cat = pd.Categorical(['b', None, 'a'], categories=['a', 'b'], ordered=True)

In [3]: cat.argsort()
Out[3]: array([1, 2, 0])

In [4]: cat[cat.argsort()]
[NaN, a, b]
categories (2, object): [a < b]

New behavior

In [61]: cat.argsort()
Out[61]: array([2, 0, 1])

In [62]: cat[cat.argsort()]
[a, b, NaN]
Categories (2, object): [a < b]

Column order is preserved when passing a list of dicts to DataFrame

Starting with Python 3.7 the key-order of dict is guaranteed. In practice, this has been true since Python 3.6. The DataFrame constructor now treats a list of dicts in the same way as it does a list of OrderedDict, i.e. preserving the order of the dicts. This change applies only when pandas is running on Python>=3.6 (GH27309).

In [63]: data = [
   ....:     {'name': 'Joe', 'state': 'NY', 'age': 18},
   ....:     {'name': 'Jane', 'state': 'KY', 'age': 19, 'hobby': 'Minecraft'},
   ....:     {'name': 'Jean', 'state': 'OK', 'age': 20, 'finances': 'good'}
   ....: ]

Previous Behavior:

The columns were lexicographically sorted previously,

In [1]: pd.DataFrame(data)
   age finances      hobby  name state
0   18      NaN        NaN   Joe    NY
1   19      NaN  Minecraft  Jane    KY
2   20     good        NaN  Jean    OK

New Behavior:

The column order now matches the insertion-order of the keys in the dict, considering all the records from top to bottom. As a consequence, the column order of the resulting DataFrame has changed compared to previous pandas verisons.

In [64]: pd.DataFrame(data)
   name state  age      hobby finances
0   Joe    NY   18        NaN      NaN
1  Jane    KY   19  Minecraft      NaN
2  Jean    OK   20        NaN     good

[3 rows x 5 columns]

Increased minimum versions for dependencies

Due to dropping support for Python 2.7, a number of optional dependencies have updated minimum versions (GH25725, GH24942, GH25752). Independently, some minimum supported versions of dependencies were updated (GH23519, GH25554). If installed, we now require:

Package Minimum Version Required
numpy 1.13.3 X
pytz 2015.4 X
python-dateutil 2.6.1 X
bottleneck 1.2.1  
numexpr 2.6.2  
pytest (dev) 4.0.2  

For optional libraries the general recommendation is to use the latest version. The following table lists the lowest version per library that is currently being tested throughout the development of pandas. Optional libraries below the lowest tested version may still work, but are not considered supported.

Package Minimum Version
beautifulsoup4 4.6.0
fastparquet 0.2.1
gcsfs 0.2.2
lxml 3.8.0
matplotlib 2.2.2
openpyxl 2.4.8
pyarrow 0.9.0
pymysql 0.7.1
pytables 3.4.2
scipy 0.19.0
sqlalchemy 1.1.4
xarray 0.8.2
xlrd 1.1.0
xlsxwriter 0.9.8
xlwt 1.2.0

See Dependencies and Optional dependencies for more.

Other API changes


Sparse subclasses

The SparseSeries and SparseDataFrame subclasses are deprecated. Their functionality is better-provided by a Series or DataFrame with sparse values.

Previous way

In [65]: df = pd.SparseDataFrame({"A": [0, 0, 1, 2]})

In [66]: df.dtypes
A    Sparse[int64, nan]
Length: 1, dtype: object

New way

In [67]: df = pd.DataFrame({"A": pd.SparseArray([0, 0, 1, 2])})

In [68]: df.dtypes
A    Sparse[int64, 0]
Length: 1, dtype: object

The memory usage of the two approaches is identical. See Migrating for more (GH19239).

msgpack format

The msgpack format is deprecated as of 0.25 and will be removed in a future version. It is recommended to use pyarrow for on-the-wire transmission of pandas objects. (GH27084)

Other deprecations

Removal of prior version deprecations/changes

Performance improvements

  • Significant speedup in SparseArray initialization that benefits most operations, fixing performance regression introduced in v0.20.0 (GH24985)
  • DataFrame.to_stata() is now faster when outputting data with any string or non-native endian columns (GH25045)
  • Improved performance of Series.searchsorted(). The speedup is especially large when the dtype is int8/int16/int32 and the searched key is within the integer bounds for the dtype (GH22034)
  • Improved performance of pandas.core.groupby.GroupBy.quantile() (GH20405)
  • Improved performance of slicing and other selected operation on a RangeIndex (GH26565, GH26617, GH26722)
  • RangeIndex now performs standard lookup without instantiating an actual hashtable, hence saving memory (GH16685)
  • Improved performance of read_csv() by faster tokenizing and faster parsing of small float numbers (GH25784)
  • Improved performance of read_csv() by faster parsing of N/A and boolean values (GH25804)
  • Improved performance of IntervalIndex.is_monotonic, IntervalIndex.is_monotonic_increasing and IntervalIndex.is_monotonic_decreasing by removing conversion to MultiIndex (GH24813)
  • Improved performance of DataFrame.to_csv() when writing datetime dtypes (GH25708)
  • Improved performance of read_csv() by much faster parsing of MM/YYYY and DD/MM/YYYY datetime formats (GH25922)
  • Improved performance of nanops for dtypes that cannot store NaNs. Speedup is particularly prominent for Series.all() and Series.any() (GH25070)
  • Improved performance of for dictionary mappers on categorical series by mapping the categories instead of mapping all values (GH23785)
  • Improved performance of IntervalIndex.intersection() (GH24813)
  • Improved performance of read_csv() by faster concatenating date columns without extra conversion to string for integer/float zero and float NaN; by faster checking the string for the possibility of being a date (GH25754)
  • Improved performance of IntervalIndex.is_unique by removing conversion to MultiIndex (GH24813)
  • Restored performance of DatetimeIndex.__iter__() by re-enabling specialized code path (GH26702)
  • Improved performance when building MultiIndex with at least one CategoricalIndex level (GH22044)
  • Improved performance by removing the need for a garbage collect when checking for SettingWithCopyWarning (GH27031)
  • For to_datetime() changed default value of cache parameter to True (GH26043)
  • Improved performance of DatetimeIndex and PeriodIndex slicing given non-unique, monotonic data (GH27136).
  • Improved performance of pd.read_json() for index-oriented data. (GH26773)
  • Improved performance of MultiIndex.shape() (GH27384).

Bug fixes



  • Bug in to_datetime() which would raise an (incorrect) ValueError when called with a date far into the future and the format argument specified instead of raising OutOfBoundsDatetime (GH23830)
  • Bug in to_datetime() which would raise InvalidIndexError: Reindexing only valid with uniquely valued Index objects when called with cache=True, with arg including at least two different elements from the set {None, numpy.nan, pandas.NaT} (GH22305)
  • Bug in DataFrame and Series where timezone aware data with dtype='datetime64[ns] was not cast to naive (GH25843)
  • Improved Timestamp type checking in various datetime functions to prevent exceptions when using a subclassed datetime (GH25851)
  • Bug in Series and DataFrame repr where np.datetime64('NaT') and np.timedelta64('NaT') with dtype=object would be represented as NaN (GH25445)
  • Bug in to_datetime() which does not replace the invalid argument with NaT when error is set to coerce (GH26122)
  • Bug in adding DateOffset with nonzero month to DatetimeIndex would raise ValueError (GH26258)
  • Bug in to_datetime() which raises unhandled OverflowError when called with mix of invalid dates and NaN values with format='%Y%m%d' and error='coerce' (GH25512)
  • Bug in isin() for datetimelike indexes; DatetimeIndex, TimedeltaIndex and PeriodIndex where the levels parameter was ignored. (GH26675)
  • Bug in to_datetime() which raises TypeError for format='%Y%m%d' when called for invalid integer dates with length >= 6 digits with errors='ignore'
  • Bug when comparing a PeriodIndex against a zero-dimensional numpy array (GH26689)
  • Bug in constructing a Series or DataFrame from a numpy datetime64 array with a non-ns unit and out-of-bound timestamps generating rubbish data, which will now correctly raise an OutOfBoundsDatetime error (GH26206).
  • Bug in date_range() with unnecessary OverflowError being raised for very large or very small dates (GH26651)
  • Bug where adding Timestamp to a np.timedelta64 object would raise instead of returning a Timestamp (GH24775)
  • Bug where comparing a zero-dimensional numpy array containing a np.datetime64 object to a Timestamp would incorrect raise TypeError (GH26916)
  • Bug in to_datetime() which would raise ValueError: Tz-aware datetime.datetime cannot be converted to datetime64 unless utc=True when called with cache=True, with arg including datetime strings with different offset (GH26097)


  • Bug in TimedeltaIndex.intersection() where for non-monotonic indices in some cases an empty Index was returned when in fact an intersection existed (GH25913)
  • Bug with comparisons between Timedelta and NaT raising TypeError (GH26039)
  • Bug when adding or subtracting a BusinessHour to a Timestamp with the resulting time landing in a following or prior day respectively (GH26381)
  • Bug when comparing a TimedeltaIndex against a zero-dimensional numpy array (GH26689)



  • Bug in to_numeric() in which large negative numbers were being improperly handled (GH24910)
  • Bug in to_numeric() in which numbers were being coerced to float, even though errors was not coerce (GH24910)
  • Bug in to_numeric() in which invalid values for errors were being allowed (GH26466)
  • Bug in format in which floating point complex numbers were not being formatted to proper display precision and trimming (GH25514)
  • Bug in error messages in DataFrame.corr() and Series.corr(). Added the possibility of using a callable. (GH25729)
  • Bug in Series.divmod() and Series.rdivmod() which would raise an (incorrect) ValueError rather than return a pair of Series objects as result (GH25557)
  • Raises a helpful exception when a non-numeric index is sent to interpolate() with methods which require numeric index. (GH21662)
  • Bug in eval() when comparing floats with scalar operators, for example: x < -0.1 (GH25928)
  • Fixed bug where casting all-boolean array to integer extension array failed (GH25211)
  • Bug in divmod with a Series object containing zeros incorrectly raising AttributeError (GH26987)
  • Inconsistency in Series floor-division (//) and divmod filling positive//zero with NaN instead of Inf (GH27321)





  • Improved exception message when calling DataFrame.iloc() with a list of non-numeric objects (GH25753).
  • Improved exception message when calling .iloc or .loc with a boolean indexer with different length (GH26658).
  • Bug in KeyError exception message when indexing a MultiIndex with a non-existant key not displaying the original key (GH27250).
  • Bug in .iloc and .loc with a boolean indexer not raising an IndexError when too few items are passed (GH26658).
  • Bug in DataFrame.loc() and Series.loc() where KeyError was not raised for a MultiIndex when the key was less than or equal to the number of levels in the MultiIndex (GH14885).
  • Bug in which DataFrame.append() produced an erroneous warning indicating that a KeyError will be thrown in the future when the data to be appended contains new columns (GH22252).
  • Bug in which DataFrame.to_csv() caused a segfault for a reindexed data frame, when the indices were single-level MultiIndex (GH26303).
  • Fixed bug where assigning a arrays.PandasArray to a pandas.core.frame.DataFrame would raise error (GH26390)
  • Allow keyword arguments for callable local reference used in the DataFrame.query() string (GH26426)
  • Fixed a KeyError when indexing a MultiIndex` level with a list containing exactly one label, which is missing (GH27148)
  • Bug which produced AttributeError on partial matching Timestamp in a MultiIndex (GH26944)
  • Bug in Categorical and CategoricalIndex with Interval values when using the in operator (__contains) with objects that are not comparable to the values in the Interval (GH23705)
  • Bug in DataFrame.loc() and DataFrame.iloc() on a DataFrame with a single timezone-aware datetime64[ns] column incorrectly returning a scalar instead of a Series (GH27110)
  • Bug in CategoricalIndex and Categorical incorrectly raising ValueError instead of TypeError when a list is passed using the in operator (__contains__) (GH21729)
  • Bug in setting a new value in a Series with a Timedelta object incorrectly casting the value to an integer (GH22717)
  • Bug in Series setting a new key (__setitem__) with a timezone-aware datetime incorrectly raising ValueError (GH12862)
  • Bug in DataFrame.iloc() when indexing with a read-only indexer (GH17192)
  • Bug in Series setting an existing tuple key (__setitem__) with timezone-aware datetime values incorrectly raising TypeError (GH20441)




  • Bug in DataFrame.to_html() where values were truncated using display options instead of outputting the full content (GH17004)
  • Fixed bug in missing text when using to_clipboard() if copying utf-16 characters in Python 3 on Windows (GH25040)
  • Bug in read_json() for orient='table' when it tries to infer dtypes by default, which is not applicable as dtypes are already defined in the JSON schema (GH21345)
  • Bug in read_json() for orient='table' and float index, as it infers index dtype by default, which is not applicable because index dtype is already defined in the JSON schema (GH25433)
  • Bug in read_json() for orient='table' and string of float column names, as it makes a column name type conversion to Timestamp, which is not applicable because column names are already defined in the JSON schema (GH25435)
  • Bug in json_normalize() for errors='ignore' where missing values in the input data, were filled in resulting DataFrame with the string "nan" instead of numpy.nan (GH25468)
  • DataFrame.to_html() now raises TypeError when using an invalid type for the classes parameter instead of AssertionError (GH25608)
  • Bug in DataFrame.to_string() and DataFrame.to_latex() that would lead to incorrect output when the header keyword is used (GH16718)
  • Bug in read_csv() not properly interpreting the UTF8 encoded filenames on Windows on Python 3.6+ (GH15086)
  • Improved performance in pandas.read_stata() and when converting columns that have missing values (GH25772)
  • Bug in DataFrame.to_html() where header numbers would ignore display options when rounding (GH17280)
  • Bug in read_hdf() where reading a table from an HDF5 file written directly with PyTables fails with a ValueError when using a sub-selection via the start or stop arguments (GH11188)
  • Bug in read_hdf() not properly closing store after a KeyError is raised (GH25766)
  • Improved the explanation for the failure when value labels are repeated in Stata dta files and suggested work-arounds (GH25772)
  • Improved pandas.read_stata() and to read incorrectly formatted 118 format files saved by Stata (GH25960)
  • Improved the col_space parameter in DataFrame.to_html() to accept a string so CSS length values can be set correctly (GH25941)
  • Fixed bug in loading objects from S3 that contain # characters in the URL (GH25945)
  • Adds use_bqstorage_api parameter to read_gbq() to speed up downloads of large data frames. This feature requires version 0.10.0 of the pandas-gbq library as well as the google-cloud-bigquery-storage and fastavro libraries. (GH26104)
  • Fixed memory leak in DataFrame.to_json() when dealing with numeric data (GH24889)
  • Bug in read_json() where date strings with Z were not converted to a UTC timezone (GH26168)
  • Added cache_dates=True parameter to read_csv(), which allows to cache unique dates when they are parsed (GH25990)
  • DataFrame.to_excel() now raises a ValueError when the caller’s dimensions exceed the limitations of Excel (GH26051)
  • Fixed bug in pandas.read_csv() where a BOM would result in incorrect parsing using engine=’python’ (GH26545)
  • read_excel() now raises a ValueError when input is of type and engine param is passed since has an engine defined (GH26566)
  • Bug while selecting from HDFStore with where='' specified (GH26610).
  • Fixed bug in DataFrame.to_excel() where custom objects (i.e. PeriodIndex) inside merged cells were not being converted into types safe for the Excel writer (GH27006)
  • Bug in read_hdf() where reading a timezone aware DatetimeIndex would raise a TypeError (GH11926)
  • Bug in to_msgpack() and read_msgpack() which would raise a ValueError rather than a FileNotFoundError for an invalid path (GH27160)
  • Fixed bug in DataFrame.to_parquet() which would raise a ValueError when the dataframe had no columns (GH27339)
  • Allow parsing of PeriodDtype columns when using read_csv() (GH26934)




  • Bug in pandas.merge() adds a string of None, if None is assigned in suffixes instead of remain the column name as-is (GH24782).
  • Bug in merge() when merging by index name would sometimes result in an incorrectly numbered index (missing index values are now assigned NA) (GH24212, GH25009)
  • to_records() now accepts dtypes to its column_dtypes parameter (GH24895)
  • Bug in concat() where order of OrderedDict (and dict in Python 3.6+) is not respected, when passed in as objs argument (GH21510)
  • Bug in pivot_table() where columns with NaN values are dropped even if dropna argument is False, when the aggfunc argument contains a list (GH22159)
  • Bug in concat() where the resulting freq of two DatetimeIndex with the same freq would be dropped (GH3232).
  • Bug in merge() where merging with equivalent Categorical dtypes was raising an error (GH22501)
  • bug in DataFrame instantiating with a dict of iterators or generators (e.g. pd.DataFrame({'A': reversed(range(3))})) raised an error (GH26349).
  • Bug in DataFrame instantiating with a range (e.g. pd.DataFrame(range(3))) raised an error (GH26342).
  • Bug in DataFrame constructor when passing non-empty tuples would cause a segmentation fault (GH25691)
  • Bug in Series.apply() failed when the series is a timezone aware DatetimeIndex (GH25959)
  • Bug in pandas.cut() where large bins could incorrectly raise an error due to an integer overflow (GH26045)
  • Bug in DataFrame.sort_index() where an error is thrown when a multi-indexed DataFrame is sorted on all levels with the initial level sorted last (GH26053)
  • Bug in Series.nlargest() treats True as smaller than False (GH26154)
  • Bug in DataFrame.pivot_table() with a IntervalIndex as pivot index would raise TypeError (GH25814)
  • Bug in which DataFrame.from_dict() ignored order of OrderedDict when orient='index' (GH8425).
  • Bug in DataFrame.transpose() where transposing a DataFrame with a timezone-aware datetime column would incorrectly raise ValueError (GH26825)
  • Bug in pivot_table() when pivoting a timezone aware column as the values would remove timezone information (GH14948)
  • Bug in merge_asof() when specifying multiple by columns where one is datetime64[ns, tz] dtype (GH26649)


  • Significant speedup in SparseArray initialization that benefits most operations, fixing performance regression introduced in v0.20.0 (GH24985)
  • Bug in SparseFrame constructor where passing None as the data would cause default_fill_value to be ignored (GH16807)
  • Bug in SparseDataFrame when adding a column in which the length of values does not match length of index, AssertionError is raised instead of raising ValueError (GH25484)
  • Introduce a better error message in Series.sparse.from_coo() so it returns a TypeError for inputs that are not coo matrices (GH26554)
  • Bug in numpy.modf() on a SparseArray. Now a tuple of SparseArray is returned (GH26946).

Build Changes

  • Fix install error with PyPy on macOS (GH26536)


  • Bug in factorize() when passing an ExtensionArray with a custom na_sentinel (GH25696).
  • Series.count() miscounts NA values in ExtensionArrays (GH26835)
  • Added Series.__array_ufunc__ to better handle NumPy ufuncs applied to Series backed by extension arrays (GH23293).
  • Keyword argument deep has been removed from ExtensionArray.copy() (GH27083)


  • Removed unused C functions from vendored UltraJSON implementation (GH26198)
  • Allow Index and RangeIndex to be passed to numpy min and max functions (GH26125)
  • Use actual class name in repr of empty objects of a Series subclass (GH27001).
  • Bug in DataFrame where passing an object array of timezone-aware datetime objects would incorrectly raise ValueError (GH13287)


Scroll To Top