What’s new in 0.23.0 (May 15, 2018)

This is a major release from 0.22.0 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:

Check the API Changes and deprecations before updating.


Starting January 1, 2019, pandas feature releases will support Python 3 only. See Dropping Python 2.7 for more.

What’s new in v0.23.0

New features

JSON read/write round-trippable with orient='table'

A DataFrame can now be written to and subsequently read back via JSON while preserving metadata through usage of the orient='table' argument (see GH18912 and GH9146). Previously, none of the available orient values guaranteed the preservation of dtypes and index names, amongst other metadata.

In [1]: 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 [2]: df
     foo bar        baz qux
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

[4 rows x 4 columns]

In [3]: df.dtypes
foo             int64
bar            object
baz    datetime64[ns]
qux          category
Length: 4, dtype: object

In [4]: df.to_json('test.json', orient='table')

In [5]: new_df = pd.read_json('test.json', orient='table')

In [6]: new_df
     foo bar        baz qux
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

[4 rows x 4 columns]

In [7]: new_df.dtypes
foo             int64
bar            object
baz    datetime64[ns]
qux          category
Length: 4, dtype: object

Please note that the string index is not supported with the round trip format, as it is used by default in write_json to indicate a missing index name.

In [8]: df.index.name = 'index'

In [9]: df.to_json('test.json', orient='table')

In [10]: new_df = pd.read_json('test.json', orient='table')

In [11]: new_df
   foo bar        baz qux
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

[4 rows x 4 columns]

In [12]: new_df.dtypes
foo             int64
bar            object
baz    datetime64[ns]
qux          category
Length: 4, dtype: object

Method .assign() accepts dependent arguments

The DataFrame.assign() now accepts dependent keyword arguments for python version later than 3.6 (see also PEP 468). Later keyword arguments may now refer to earlier ones if the argument is a callable. See the documentation here (GH14207)

In [13]: df = pd.DataFrame({'A': [1, 2, 3]})

In [14]: df
0  1
1  2
2  3

[3 rows x 1 columns]

In [15]: df.assign(B=df.A, C=lambda x: x['A'] + x['B'])
   A  B  C
0  1  1  2
1  2  2  4
2  3  3  6

[3 rows x 3 columns]


This may subtly change the behavior of your code when you’re using .assign() to update an existing column. Previously, callables referring to other variables being updated would get the “old” values

Previous behavior:

In [2]: df = pd.DataFrame({"A": [1, 2, 3]})

In [3]: df.assign(A=lambda df: df.A + 1, C=lambda df: df.A * -1)
   A  C
0  2 -1
1  3 -2
2  4 -3

New behavior:

In [16]: df.assign(A=df.A + 1, C=lambda df: df.A * -1)
   A  C
0  2 -2
1  3 -3
2  4 -4

[3 rows x 2 columns]

Merging on a combination of columns and index levels

Strings passed to DataFrame.merge() as the on, left_on, and right_on parameters may now refer to either column names or index level names. This enables merging DataFrame instances on a combination of index levels and columns without resetting indexes. See the Merge on columns and levels documentation section. (GH14355)

In [17]: left_index = pd.Index(['K0', 'K0', 'K1', 'K2'], name='key1')

In [18]: left = pd.DataFrame({'A': ['A0', 'A1', 'A2', 'A3'],
   ....:                      'B': ['B0', 'B1', 'B2', 'B3'],
   ....:                      'key2': ['K0', 'K1', 'K0', 'K1']},
   ....:                     index=left_index)

In [19]: right_index = pd.Index(['K0', 'K1', 'K2', 'K2'], name='key1')

In [20]: right = pd.DataFrame({'C': ['C0', 'C1', 'C2', 'C3'],
   ....:                       'D': ['D0', 'D1', 'D2', 'D3'],
   ....:                       'key2': ['K0', 'K0', 'K0', 'K1']},
   ....:                      index=right_index)

In [21]: left.merge(right, on=['key1', 'key2'])
       A   B key2   C   D
K0    A0  B0   K0  C0  D0
K1    A2  B2   K0  C1  D1
K2    A3  B3   K1  C3  D3

[3 rows x 5 columns]

Sorting by a combination of columns and index levels

Strings passed to DataFrame.sort_values() as the by parameter may now refer to either column names or index level names. This enables sorting DataFrame instances by a combination of index levels and columns without resetting indexes. See the Sorting by Indexes and Values documentation section. (GH14353)

# Build MultiIndex
In [22]: idx = pd.MultiIndex.from_tuples([('a', 1), ('a', 2), ('a', 2),
   ....:                                  ('b', 2), ('b', 1), ('b', 1)])

In [23]: idx.names = ['first', 'second']

# Build DataFrame
In [24]: df_multi = pd.DataFrame({'A': np.arange(6, 0, -1)},
   ....:                         index=idx)

In [25]: df_multi
first second   
a     1       6
      2       5
      2       4
b     2       3
      1       2
      1       1

[6 rows x 1 columns]

# Sort by 'second' (index) and 'A' (column)
In [26]: df_multi.sort_values(by=['second', 'A'])
first second   
b     1       1
      1       2
a     1       6
b     2       3
a     2       4
      2       5

[6 rows x 1 columns]

Extending pandas with custom types (experimental)

pandas now supports storing array-like objects that aren’t necessarily 1-D NumPy arrays as columns in a DataFrame or values in a Series. This allows third-party libraries to implement extensions to NumPy’s types, similar to how pandas implemented categoricals, datetimes with timezones, periods, and intervals.

As a demonstration, we’ll use cyberpandas, which provides an IPArray type for storing ip addresses.

In [1]: from cyberpandas import IPArray

In [2]: values = IPArray([
   ...:     0,
   ...:     3232235777,
   ...:     42540766452641154071740215577757643572
   ...: ])

IPArray isn’t a normal 1-D NumPy array, but because it’s a pandas ExtensionArray, it can be stored properly inside pandas’ containers.

In [3]: ser = pd.Series(values)

In [4]: ser
2    2001:db8:85a3::8a2e:370:7334
dtype: ip

Notice that the dtype is ip. The missing value semantics of the underlying array are respected:

In [5]: ser.isna()
0     True
1    False
2    False
dtype: bool

For more, see the extension types documentation. If you build an extension array, publicize it on our ecosystem page.

New observed keyword for excluding unobserved categories in GroupBy

Grouping by a categorical includes the unobserved categories in the output. When grouping by multiple categorical columns, this means you get the cartesian product of all the categories, including combinations where there are no observations, which can result in a large number of groups. We have added a keyword observed to control this behavior, it defaults to observed=False for backward-compatibility. (GH14942, GH8138, GH15217, GH17594, GH8669, GH20583, GH20902)

In [27]: cat1 = pd.Categorical(["a", "a", "b", "b"],
   ....:                       categories=["a", "b", "z"], ordered=True)

In [28]: cat2 = pd.Categorical(["c", "d", "c", "d"],
   ....:                       categories=["c", "d", "y"], ordered=True)

In [29]: df = pd.DataFrame({"A": cat1, "B": cat2, "values": [1, 2, 3, 4]})

In [30]: df['C'] = ['foo', 'bar'] * 2

In [31]: df
   A  B  values    C
0  a  c       1  foo
1  a  d       2  bar
2  b  c       3  foo
3  b  d       4  bar

[4 rows x 4 columns]

To show all values, the previous behavior:

In [32]: df.groupby(['A', 'B', 'C'], observed=False).count()
A B C          
a c bar       0
    foo       1
  d bar       1
    foo       0
  y bar       0
...         ...
z c foo       0
  d bar       0
    foo       0
  y bar       0
    foo       0

[18 rows x 1 columns]

To show only observed values:

In [33]: df.groupby(['A', 'B', 'C'], observed=True).count()
A B C          
a c foo       1
  d bar       1
b c foo       1
  d bar       1

[4 rows x 1 columns]

For pivoting operations, this behavior is already controlled by the dropna keyword:

In [34]: cat1 = pd.Categorical(["a", "a", "b", "b"],
   ....:                       categories=["a", "b", "z"], ordered=True)

In [35]: cat2 = pd.Categorical(["c", "d", "c", "d"],
   ....:                       categories=["c", "d", "y"], ordered=True)

In [36]: df = pd.DataFrame({"A": cat1, "B": cat2, "values": [1, 2, 3, 4]})

In [37]: df
   A  B  values
0  a  c       1
1  a  d       2
2  b  c       3
3  b  d       4

[4 rows x 3 columns]
In [38]: pd.pivot_table(df, values='values', index=['A', 'B'],
   ....:                dropna=True)
A B        
a c       1
  d       2
b c       3
  d       4

[4 rows x 1 columns]

In [39]: pd.pivot_table(df, values='values', index=['A', 'B'],
   ....:                dropna=False)
A B        
a c     1.0
  d     2.0
  y     NaN
b c     3.0
  d     4.0
  y     NaN
z c     NaN
  d     NaN
  y     NaN

[9 rows x 1 columns]

Rolling/Expanding.apply() accepts raw=False to pass a Series to the function

Series.rolling().apply(), DataFrame.rolling().apply(), Series.expanding().apply(), and DataFrame.expanding().apply() have gained a raw=None parameter. This is similar to DataFame.apply(). This parameter, if True allows one to send a np.ndarray to the applied function. If False a Series will be passed. The default is None, which preserves backward compatibility, so this will default to True, sending an np.ndarray. In a future version the default will be changed to False, sending a Series. (GH5071, GH20584)

In [40]: s = pd.Series(np.arange(5), np.arange(5) + 1)

In [41]: s
1    0
2    1
3    2
4    3
5    4
Length: 5, dtype: int64

Pass a Series:

In [42]: s.rolling(2, min_periods=1).apply(lambda x: x.iloc[-1], raw=False)
1    0.0
2    1.0
3    2.0
4    3.0
5    4.0
Length: 5, dtype: float64

Mimic the original behavior of passing a ndarray:

In [43]: s.rolling(2, min_periods=1).apply(lambda x: x[-1], raw=True)
1    0.0
2    1.0
3    2.0
4    3.0
5    4.0
Length: 5, dtype: float64

DataFrame.interpolate has gained the limit_area kwarg

DataFrame.interpolate() has gained a limit_area parameter to allow further control of which NaN s are replaced. Use limit_area='inside' to fill only NaNs surrounded by valid values or use limit_area='outside' to fill only NaN s outside the existing valid values while preserving those inside. (GH16284) See the full documentation here.

In [44]: ser = pd.Series([np.nan, np.nan, 5, np.nan, np.nan,
   ....:                  np.nan, 13, np.nan, np.nan])

In [45]: ser
0     NaN
1     NaN
2     5.0
3     NaN
4     NaN
5     NaN
6    13.0
7     NaN
8     NaN
Length: 9, dtype: float64

Fill one consecutive inside value in both directions

In [46]: ser.interpolate(limit_direction='both', limit_area='inside', limit=1)
0     NaN
1     NaN
2     5.0
3     7.0
4     NaN
5    11.0
6    13.0
7     NaN
8     NaN
Length: 9, dtype: float64

Fill all consecutive outside values backward

In [47]: ser.interpolate(limit_direction='backward', limit_area='outside')
0     5.0
1     5.0
2     5.0
3     NaN
4     NaN
5     NaN
6    13.0
7     NaN
8     NaN
Length: 9, dtype: float64

Fill all consecutive outside values in both directions

In [48]: ser.interpolate(limit_direction='both', limit_area='outside')
0     5.0
1     5.0
2     5.0
3     NaN
4     NaN
5     NaN
6    13.0
7    13.0
8    13.0
Length: 9, dtype: float64

Function get_dummies now supports dtype argument

The get_dummies() now accepts a dtype argument, which specifies a dtype for the new columns. The default remains uint8. (GH18330)

In [49]: df = pd.DataFrame({'a': [1, 2], 'b': [3, 4], 'c': [5, 6]})

In [50]: pd.get_dummies(df, columns=['c']).dtypes
a      int64
b      int64
c_5    uint8
c_6    uint8
Length: 4, dtype: object

In [51]: pd.get_dummies(df, columns=['c'], dtype=bool).dtypes
a      int64
b      int64
c_5     bool
c_6     bool
Length: 4, dtype: object

Timedelta mod method

mod (%) and divmod operations are now defined on Timedelta objects when operating with either timedelta-like or with numeric arguments. See the documentation here. (GH19365)

In [52]: td = pd.Timedelta(hours=37)

In [53]: td % pd.Timedelta(minutes=45)
Out[53]: Timedelta('0 days 00:15:00')

Method .rank() handles inf values when NaN are present

In previous versions, .rank() would assign inf elements NaN as their ranks. Now ranks are calculated properly. (GH6945)

In [54]: s = pd.Series([-np.inf, 0, 1, np.nan, np.inf])

In [55]: s
0   -inf
1    0.0
2    1.0
3    NaN
4    inf
Length: 5, dtype: float64

Previous behavior:

In [11]: s.rank()
0    1.0
1    2.0
2    3.0
3    NaN
4    NaN
dtype: float64

Current behavior:

In [56]: s.rank()
0    1.0
1    2.0
2    3.0
3    NaN
4    4.0
Length: 5, dtype: float64

Furthermore, previously if you rank inf or -inf values together with NaN values, the calculation won’t distinguish NaN from infinity when using ‘top’ or ‘bottom’ argument.

In [57]: s = pd.Series([np.nan, np.nan, -np.inf, -np.inf])

In [58]: s
0    NaN
1    NaN
2   -inf
3   -inf
Length: 4, dtype: float64

Previous behavior:

In [15]: s.rank(na_option='top')
0    2.5
1    2.5
2    2.5
3    2.5
dtype: float64

Current behavior:

In [59]: s.rank(na_option='top')
0    1.5
1    1.5
2    3.5
3    3.5
Length: 4, dtype: float64

These bugs were squashed:

  • Bug in DataFrame.rank() and Series.rank() when method='dense' and pct=True in which percentile ranks were not being used with the number of distinct observations (GH15630)

  • Bug in Series.rank() and DataFrame.rank() when ascending='False' failed to return correct ranks for infinity if NaN were present (GH19538)

  • Bug in DataFrameGroupBy.rank() where ranks were incorrect when both infinity and NaN were present (GH20561)

Series.str.cat has gained the join kwarg

Previously, Series.str.cat() did not – in contrast to most of pandas – align Series on their index before concatenation (see GH18657). The method has now gained a keyword join to control the manner of alignment, see examples below and here.

In v.0.23 join will default to None (meaning no alignment), but this default will change to 'left' in a future version of pandas.

In [60]: s = pd.Series(['a', 'b', 'c', 'd'])

In [61]: t = pd.Series(['b', 'd', 'e', 'c'], index=[1, 3, 4, 2])

In [62]: s.str.cat(t)
0    NaN
1     bb
2     cc
3     dd
Length: 4, dtype: object

In [63]: s.str.cat(t, join='left', na_rep='-')
0    a-
1    bb
2    cc
3    dd
Length: 4, dtype: object

Furthermore, Series.str.cat() now works for CategoricalIndex as well (previously raised a ValueError; see GH20842).

DataFrame.astype performs column-wise conversion to Categorical

DataFrame.astype() can now perform column-wise conversion to Categorical by supplying the string 'category' or a CategoricalDtype. Previously, attempting this would raise a NotImplementedError. See the Object creation section of the documentation for more details and examples. (GH12860, GH18099)

Supplying the string 'category' performs column-wise conversion, with only labels appearing in a given column set as categories:

In [64]: df = pd.DataFrame({'A': list('abca'), 'B': list('bccd')})

In [65]: df = df.astype('category')

In [66]: df['A'].dtype
Out[66]: CategoricalDtype(categories=['a', 'b', 'c'], ordered=False)

In [67]: df['B'].dtype
Out[67]: CategoricalDtype(categories=['b', 'c', 'd'], ordered=False)

Supplying a CategoricalDtype will make the categories in each column consistent with the supplied dtype:

In [68]: from pandas.api.types import CategoricalDtype

In [69]: df = pd.DataFrame({'A': list('abca'), 'B': list('bccd')})

In [70]: cdt = CategoricalDtype(categories=list('abcd'), ordered=True)

In [71]: df = df.astype(cdt)

In [72]: df['A'].dtype
Out[72]: CategoricalDtype(categories=['a', 'b', 'c', 'd'], ordered=True)

In [73]: df['B'].dtype
Out[73]: CategoricalDtype(categories=['a', 'b', 'c', 'd'], ordered=True)

Other enhancements

Backwards incompatible API changes

Dependencies have increased minimum versions

We have updated our minimum supported versions of dependencies (GH15184). If installed, we now require:


Minimum Version
















Instantiation from dicts preserves dict insertion order for Python 3.6+

Until Python 3.6, dicts in Python had no formally defined ordering. For Python version 3.6 and later, dicts are ordered by insertion order, see PEP 468. pandas will use the dict’s insertion order, when creating a Series or DataFrame from a dict and you’re using Python version 3.6 or higher. (GH19884)

Previous behavior (and current behavior if on Python < 3.6):

In [16]: pd.Series({'Income': 2000,
   ....:            'Expenses': -1500,
   ....:            'Taxes': -200,
   ....:            'Net result': 300})
Expenses     -1500
Income        2000
Net result     300
Taxes         -200
dtype: int64

Note the Series above is ordered alphabetically by the index values.

New behavior (for Python >= 3.6):

In [74]: pd.Series({'Income': 2000,
   ....:            'Expenses': -1500,
   ....:            'Taxes': -200,
   ....:            'Net result': 300})
Income        2000
Expenses     -1500
Taxes         -200
Net result     300
Length: 4, dtype: int64

Notice that the Series is now ordered by insertion order. This new behavior is used for all relevant pandas types (Series, DataFrame, SparseSeries and SparseDataFrame).

If you wish to retain the old behavior while using Python >= 3.6, you can use .sort_index():

In [75]: pd.Series({'Income': 2000,
   ....:            'Expenses': -1500,
   ....:            'Taxes': -200,
   ....:            'Net result': 300}).sort_index()
Expenses     -1500
Income        2000
Net result     300
Taxes         -200
Length: 4, dtype: int64

Deprecate Panel

Panel was deprecated in the 0.20.x release, showing as a DeprecationWarning. Using Panel will now show a FutureWarning. The recommended way to represent 3-D data are with a MultiIndex on a DataFrame via the to_frame() or with the xarray package. pandas provides a to_xarray() method to automate this conversion (GH13563, GH18324).

In [75]: import pandas._testing as tm

In [76]: p = tm.makePanel()

In [77]: p
<class 'pandas.core.panel.Panel'>
Dimensions: 3 (items) x 3 (major_axis) x 4 (minor_axis)
Items axis: ItemA to ItemC
Major_axis axis: 2000-01-03 00:00:00 to 2000-01-05 00:00:00
Minor_axis axis: A to D

Convert to a MultiIndex DataFrame

In [78]: p.to_frame()
                     ItemA     ItemB     ItemC
major      minor
2000-01-03 A      0.469112  0.721555  0.404705
           B     -1.135632  0.271860 -1.039268
           C      0.119209  0.276232 -1.344312
           D     -2.104569  0.113648 -0.109050
2000-01-04 A     -0.282863 -0.706771  0.577046
           B      1.212112 -0.424972 -0.370647
           C     -1.044236 -1.087401  0.844885
           D     -0.494929 -1.478427  1.643563
2000-01-05 A     -1.509059 -1.039575 -1.715002
           B     -0.173215  0.567020 -1.157892
           C     -0.861849 -0.673690  1.075770
           D      1.071804  0.524988 -1.469388

[12 rows x 3 columns]

Convert to an xarray DataArray

In [79]: p.to_xarray()
<xarray.DataArray (items: 3, major_axis: 3, minor_axis: 4)>
array([[[ 0.469112, -1.135632,  0.119209, -2.104569],
        [-0.282863,  1.212112, -1.044236, -0.494929],
        [-1.509059, -0.173215, -0.861849,  1.071804]],

       [[ 0.721555,  0.27186 ,  0.276232,  0.113648],
        [-0.706771, -0.424972, -1.087401, -1.478427],
        [-1.039575,  0.56702 , -0.67369 ,  0.524988]],

       [[ 0.404705, -1.039268, -1.344312, -0.10905 ],
        [ 0.577046, -0.370647,  0.844885,  1.643563],
        [-1.715002, -1.157892,  1.07577 , -1.469388]]])
  * items       (items) object 'ItemA' 'ItemB' 'ItemC'
  * major_axis  (major_axis) datetime64[ns] 2000-01-03 2000-01-04 2000-01-05
  * minor_axis  (minor_axis) object 'A' 'B' 'C' 'D'

pandas.core.common removals

The following error & warning messages are removed from pandas.core.common (GH13634, GH19769):

  • PerformanceWarning

  • UnsupportedFunctionCall

  • UnsortedIndexError

  • AbstractMethodError

These are available from import from pandas.errors (since 0.19.0).

Changes to make output of DataFrame.apply consistent

DataFrame.apply() was inconsistent when applying an arbitrary user-defined-function that returned a list-like with axis=1. Several bugs and inconsistencies are resolved. If the applied function returns a Series, then pandas will return a DataFrame; otherwise a Series will be returned, this includes the case where a list-like (e.g. tuple or list is returned) (GH16353, GH17437, GH17970, GH17348, GH17892, GH18573, GH17602, GH18775, GH18901, GH18919).

In [76]: df = pd.DataFrame(np.tile(np.arange(3), 6).reshape(6, -1) + 1,
   ....:                   columns=['A', 'B', 'C'])

In [77]: df
   A  B  C
0  1  2  3
1  1  2  3
2  1  2  3
3  1  2  3
4  1  2  3
5  1  2  3

[6 rows x 3 columns]

Previous behavior: if the returned shape happened to match the length of original columns, this would return a DataFrame. If the return shape did not match, a Series with lists was returned.

In [3]: df.apply(lambda x: [1, 2, 3], axis=1)
   A  B  C
0  1  2  3
1  1  2  3
2  1  2  3
3  1  2  3
4  1  2  3
5  1  2  3

In [4]: df.apply(lambda x: [1, 2], axis=1)
0    [1, 2]
1    [1, 2]
2    [1, 2]
3    [1, 2]
4    [1, 2]
5    [1, 2]
dtype: object

New behavior: When the applied function returns a list-like, this will now always return a Series.

In [78]: df.apply(lambda x: [1, 2, 3], axis=1)
0    [1, 2, 3]
1    [1, 2, 3]
2    [1, 2, 3]
3    [1, 2, 3]
4    [1, 2, 3]
5    [1, 2, 3]
Length: 6, dtype: object

In [79]: df.apply(lambda x: [1, 2], axis=1)
0    [1, 2]
1    [1, 2]
2    [1, 2]
3    [1, 2]
4    [1, 2]
5    [1, 2]
Length: 6, dtype: object

To have expanded columns, you can use result_type='expand'

In [80]: df.apply(lambda x: [1, 2, 3], axis=1, result_type='expand')
   0  1  2
0  1  2  3
1  1  2  3
2  1  2  3
3  1  2  3
4  1  2  3
5  1  2  3

[6 rows x 3 columns]

To broadcast the result across the original columns (the old behaviour for list-likes of the correct length), you can use result_type='broadcast'. The shape must match the original columns.

In [81]: df.apply(lambda x: [1, 2, 3], axis=1, result_type='broadcast')
   A  B  C
0  1  2  3
1  1  2  3
2  1  2  3
3  1  2  3
4  1  2  3
5  1  2  3

[6 rows x 3 columns]

Returning a Series allows one to control the exact return structure and column names:

In [82]: df.apply(lambda x: pd.Series([1, 2, 3], index=['D', 'E', 'F']), axis=1)
   D  E  F
0  1  2  3
1  1  2  3
2  1  2  3
3  1  2  3
4  1  2  3
5  1  2  3

[6 rows x 3 columns]

Concatenation will no longer sort

In a future version of pandas pandas.concat() will no longer sort the non-concatenation axis when it is not already aligned. The current behavior is the same as the previous (sorting), but now a warning is issued when sort is not specified and the non-concatenation axis is not aligned (GH4588).

In [83]: df1 = pd.DataFrame({"a": [1, 2], "b": [1, 2]}, columns=['b', 'a'])

In [84]: df2 = pd.DataFrame({"a": [4, 5]})

In [85]: pd.concat([df1, df2])
     b  a
0  1.0  1
1  2.0  2
0  NaN  4
1  NaN  5

[4 rows x 2 columns]

To keep the previous behavior (sorting) and silence the warning, pass sort=True

In [86]: pd.concat([df1, df2], sort=True)
   a    b
0  1  1.0
1  2  2.0
0  4  NaN
1  5  NaN

[4 rows x 2 columns]

To accept the future behavior (no sorting), pass sort=False

Note that this change also applies to DataFrame.append(), which has also received a sort keyword for controlling this behavior.

Build changes

  • Building pandas for development now requires cython >= 0.24 (GH18613)

  • Building from source now explicitly requires setuptools in setup.py (GH18113)

  • Updated conda recipe to be in compliance with conda-build 3.0+ (GH18002)

Index division by zero fills correctly

Division operations on Index and subclasses will now fill division of positive numbers by zero with np.inf, division of negative numbers by zero with -np.inf and 0 / 0 with np.nan. This matches existing Series behavior. (GH19322, GH19347)

Previous behavior:

In [6]: index = pd.Int64Index([-1, 0, 1])

In [7]: index / 0
Out[7]: Int64Index([0, 0, 0], dtype='int64')

# Previous behavior yielded different results depending on the type of zero in the divisor
In [8]: index / 0.0
Out[8]: Float64Index([-inf, nan, inf], dtype='float64')

In [9]: index = pd.UInt64Index([0, 1])

In [10]: index / np.array([0, 0], dtype=np.uint64)
Out[10]: UInt64Index([0, 0], dtype='uint64')

In [11]: pd.RangeIndex(1, 5) / 0
ZeroDivisionError: integer division or modulo by zero

Current behavior:

In [87]: index = pd.Int64Index([-1, 0, 1])

# division by zero gives -infinity where negative,
# +infinity where positive, and NaN for 0 / 0
In [88]: index / 0
Out[88]: Float64Index([-inf, nan, inf], dtype='float64')

# The result of division by zero should not depend on
# whether the zero is int or float
In [89]: index / 0.0
Out[89]: Float64Index([-inf, nan, inf], dtype='float64')

In [90]: index = pd.UInt64Index([0, 1])

In [91]: index / np.array([0, 0], dtype=np.uint64)
Out[91]: Float64Index([nan, inf], dtype='float64')

In [92]: pd.RangeIndex(1, 5) / 0
Out[92]: Float64Index([inf, inf, inf, inf], dtype='float64')

Extraction of matching patterns from strings

By default, extracting matching patterns from strings with str.extract() used to return a Series if a single group was being extracted (a DataFrame if more than one group was extracted). As of pandas 0.23.0 str.extract() always returns a DataFrame, unless expand is set to False. Finally, None was an accepted value for the expand parameter (which was equivalent to False), but now raises a ValueError. (GH11386)

Previous behavior:

In [1]: s = pd.Series(['number 10', '12 eggs'])

In [2]: extracted = s.str.extract(r'.*(\d\d).*')

In [3]: extracted
Out [3]:
0    10
1    12
dtype: object

In [4]: type(extracted)
Out [4]:

New behavior:

In [93]: s = pd.Series(['number 10', '12 eggs'])

In [94]: extracted = s.str.extract(r'.*(\d\d).*')

In [95]: extracted
0  10
1  12

[2 rows x 1 columns]

In [96]: type(extracted)
Out[96]: pandas.core.frame.DataFrame

To restore previous behavior, simply set expand to False:

In [97]: s = pd.Series(['number 10', '12 eggs'])

In [98]: extracted = s.str.extract(r'.*(\d\d).*', expand=False)

In [99]: extracted
0    10
1    12
Length: 2, dtype: object

In [100]: type(extracted)
Out[100]: pandas.core.series.Series

Default value for the ordered parameter of CategoricalDtype

The default value of the ordered parameter for CategoricalDtype has changed from False to None to allow updating of categories without impacting ordered. Behavior should remain consistent for downstream objects, such as Categorical (GH18790)

In previous versions, the default value for the ordered parameter was False. This could potentially lead to the ordered parameter unintentionally being changed from True to False when users attempt to update categories if ordered is not explicitly specified, as it would silently default to False. The new behavior for ordered=None is to retain the existing value of ordered.

New behavior:

In [2]: from pandas.api.types import CategoricalDtype

In [3]: cat = pd.Categorical(list('abcaba'), ordered=True, categories=list('cba'))

In [4]: cat
[a, b, c, a, b, a]
Categories (3, object): [c < b < a]

In [5]: cdt = CategoricalDtype(categories=list('cbad'))

In [6]: cat.astype(cdt)
[a, b, c, a, b, a]
Categories (4, object): [c < b < a < d]

Notice in the example above that the converted Categorical has retained ordered=True. Had the default value for ordered remained as False, the converted Categorical would have become unordered, despite ordered=False never being explicitly specified. To change the value of ordered, explicitly pass it to the new dtype, e.g. CategoricalDtype(categories=list('cbad'), ordered=False).

Note that the unintentional conversion of ordered discussed above did not arise in previous versions due to separate bugs that prevented astype from doing any type of category to category conversion (GH10696, GH18593). These bugs have been fixed in this release, and motivated changing the default value of ordered.

Better pretty-printing of DataFrames in a terminal

Previously, the default value for the maximum number of columns was pd.options.display.max_columns=20. This meant that relatively wide data frames would not fit within the terminal width, and pandas would introduce line breaks to display these 20 columns. This resulted in an output that was relatively difficult to read:


If Python runs in a terminal, the maximum number of columns is now determined automatically so that the printed data frame fits within the current terminal width (pd.options.display.max_columns=0) (GH17023). If Python runs as a Jupyter kernel (such as the Jupyter QtConsole or a Jupyter notebook, as well as in many IDEs), this value cannot be inferred automatically and is thus set to 20 as in previous versions. In a terminal, this results in a much nicer output:


Note that if you don’t like the new default, you can always set this option yourself. To revert to the old setting, you can run this line:

pd.options.display.max_columns = 20

Datetimelike API changes

  • The default Timedelta constructor now accepts an ISO 8601 Duration string as an argument (GH19040)

  • Subtracting NaT from a Series with dtype='datetime64[ns]' returns a Series with dtype='timedelta64[ns]' instead of dtype='datetime64[ns]' (GH18808)

  • Addition or subtraction of NaT from TimedeltaIndex will return TimedeltaIndex instead of DatetimeIndex (GH19124)

  • DatetimeIndex.shift() and TimedeltaIndex.shift() will now raise NullFrequencyError (which subclasses ValueError, which was raised in older versions) when the index object frequency is None (GH19147)

  • Addition and subtraction of NaN from a Series with dtype='timedelta64[ns]' will raise a TypeError instead of treating the NaN as NaT (GH19274)

  • NaT division with datetime.timedelta will now return NaN instead of raising (GH17876)

  • Operations between a Series with dtype dtype='datetime64[ns]' and a PeriodIndex will correctly raises TypeError (GH18850)

  • Subtraction of Series with timezone-aware dtype='datetime64[ns]' with mismatched timezones will raise TypeError instead of ValueError (GH18817)

  • Timestamp will no longer silently ignore unused or invalid tz or tzinfo keyword arguments (GH17690)

  • Timestamp will no longer silently ignore invalid freq arguments (GH5168)

  • CacheableOffset and WeekDay are no longer available in the pandas.tseries.offsets module (GH17830)

  • pandas.tseries.frequencies.get_freq_group() and pandas.tseries.frequencies.DAYS are removed from the public API (GH18034)

  • Series.truncate() and DataFrame.truncate() will raise a ValueError if the index is not sorted instead of an unhelpful KeyError (GH17935)

  • Series.first and DataFrame.first will now raise a TypeError rather than NotImplementedError when index is not a DatetimeIndex (GH20725).

  • Series.last and DataFrame.last will now raise a TypeError rather than NotImplementedError when index is not a DatetimeIndex (GH20725).

  • 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, GH18226).

  • pandas.merge() provides a more informative error message when trying to merge on timezone-aware and timezone-naive columns (GH15800)

  • For DatetimeIndex and TimedeltaIndex with freq=None, addition or subtraction of integer-dtyped array or Index will raise NullFrequencyError instead of TypeError (GH19895)

  • Timestamp constructor now accepts a nanosecond keyword or positional argument (GH18898)

  • DatetimeIndex will now raise an AttributeError when the tz attribute is set after instantiation (GH3746)

  • DatetimeIndex with a pytz timezone will now return a consistent pytz timezone (GH18595)

Other API changes

  • Series.astype() and Index.astype() with an incompatible dtype will now raise a TypeError rather than a ValueError (GH18231)

  • Series construction with an object dtyped tz-aware datetime and dtype=object specified, will now return an object dtyped Series, previously this would infer the datetime dtype (GH18231)

  • A Series of dtype=category constructed from an empty dict will now have categories of dtype=object rather than dtype=float64, consistently with the case in which an empty list is passed (GH18515)

  • All-NaN levels in a MultiIndex are now assigned float rather than object dtype, promoting consistency with Index (GH17929).

  • Levels names of a MultiIndex (when not None) are now required to be unique: trying to create a MultiIndex with repeated names will raise a ValueError (GH18872)

  • Both construction and renaming of Index/MultiIndex with non-hashable name/names will now raise TypeError (GH20527)

  • Index.map() can now accept Series and dictionary input objects (GH12756, GH18482, GH18509).

  • DataFrame.unstack() will now default to filling with np.nan for object columns. (GH12815)

  • IntervalIndex constructor will raise if the closed parameter conflicts with how the input data is inferred to be closed (GH18421)

  • Inserting missing values into indexes will work for all types of indexes and automatically insert the correct type of missing value (NaN, NaT, etc.) regardless of the type passed in (GH18295)

  • When created with duplicate labels, MultiIndex now raises a ValueError. (GH17464)

  • Series.fillna() now raises a TypeError instead of a ValueError when passed a list, tuple or DataFrame as a value (GH18293)

  • pandas.DataFrame.merge() no longer casts a float column to object when merging on int and float columns (GH16572)

  • pandas.merge() now raises a ValueError when trying to merge on incompatible data types (GH9780)

  • The default NA value for UInt64Index has changed from 0 to NaN, which impacts methods that mask with NA, such as UInt64Index.where() (GH18398)

  • Refactored setup.py to use find_packages instead of explicitly listing out all subpackages (GH18535)

  • Rearranged the order of keyword arguments in read_excel() to align with read_csv() (GH16672)

  • wide_to_long() previously kept numeric-like suffixes as object dtype. Now they are cast to numeric if possible (GH17627)

  • In read_excel(), the comment argument is now exposed as a named parameter (GH18735)

  • Rearranged the order of keyword arguments in read_excel() to align with read_csv() (GH16672)

  • The options html.border and mode.use_inf_as_null were deprecated in prior versions, these will now show FutureWarning rather than a DeprecationWarning (GH19003)

  • IntervalIndex and IntervalDtype no longer support categorical, object, and string subtypes (GH19016)

  • IntervalDtype now returns True when compared against 'interval' regardless of subtype, and IntervalDtype.name now returns 'interval' regardless of subtype (GH18980)

  • KeyError now raises instead of ValueError in drop(), drop(), drop(), drop() when dropping a non-existent element in an axis with duplicates (GH19186)

  • Series.to_csv() now accepts a compression argument that works in the same way as the compression argument in DataFrame.to_csv() (GH18958)

  • Set operations (union, difference…) on IntervalIndex with incompatible index types will now raise a TypeError rather than a ValueError (GH19329)

  • DateOffset objects render more simply, e.g. <DateOffset: days=1> instead of <DateOffset: kwds={'days': 1}> (GH19403)

  • Categorical.fillna now validates its value and method keyword arguments. It now raises when both or none are specified, matching the behavior of Series.fillna() (GH19682)

  • pd.to_datetime('today') now returns a datetime, consistent with pd.Timestamp('today'); previously pd.to_datetime('today') returned a .normalized() datetime (GH19935)

  • Series.str.replace() now takes an optional regex keyword which, when set to False, uses literal string replacement rather than regex replacement (GH16808)

  • DatetimeIndex.strftime() and PeriodIndex.strftime() now return an Index instead of a numpy array to be consistent with similar accessors (GH20127)

  • Constructing a Series from a list of length 1 no longer broadcasts this list when a longer index is specified (GH19714, GH20391).

  • DataFrame.to_dict() with orient='index' no longer casts int columns to float for a DataFrame with only int and float columns (GH18580)

  • A user-defined-function that is passed to Series.rolling().aggregate(), DataFrame.rolling().aggregate(), or its expanding cousins, will now always be passed a Series, rather than a np.array; .apply() only has the raw keyword, see here. This is consistent with the signatures of .aggregate() across pandas (GH20584)

  • Rolling and Expanding types raise NotImplementedError upon iteration (GH11704).


  • Series.from_array and SparseSeries.from_array are deprecated. Use the normal constructor Series(..) and SparseSeries(..) instead (GH18213).

  • DataFrame.as_matrix is deprecated. Use DataFrame.values instead (GH18458).

  • Series.asobject, DatetimeIndex.asobject, PeriodIndex.asobject and TimeDeltaIndex.asobject have been deprecated. Use .astype(object) instead (GH18572)

  • Grouping by a tuple of keys now emits a FutureWarning and is deprecated. In the future, a tuple passed to 'by' will always refer to a single key that is the actual tuple, instead of treating the tuple as multiple keys. To retain the previous behavior, use a list instead of a tuple (GH18314)

  • Series.valid is deprecated. Use Series.dropna() instead (GH18800).

  • read_excel() has deprecated the skip_footer parameter. Use skipfooter instead (GH18836)

  • ExcelFile.parse() has deprecated sheetname in favor of sheet_name for consistency with read_excel() (GH20920).

  • The is_copy attribute is deprecated and will be removed in a future version (GH18801).

  • IntervalIndex.from_intervals is deprecated in favor of the IntervalIndex constructor (GH19263)

  • DataFrame.from_items is deprecated. Use DataFrame.from_dict() instead, or DataFrame.from_dict(OrderedDict()) if you wish to preserve the key order (GH17320, GH17312)

  • Indexing a MultiIndex or a FloatIndex with a list containing some missing keys will now show a FutureWarning, which is consistent with other types of indexes (GH17758).

  • The broadcast parameter of .apply() is deprecated in favor of result_type='broadcast' (GH18577)

  • The reduce parameter of .apply() is deprecated in favor of result_type='reduce' (GH18577)

  • The order parameter of factorize() is deprecated and will be removed in a future release (GH19727)

  • Timestamp.weekday_name, DatetimeIndex.weekday_name, and Series.dt.weekday_name are deprecated in favor of Timestamp.day_name(), DatetimeIndex.day_name(), and Series.dt.day_name() (GH12806)

  • pandas.tseries.plotting.tsplot is deprecated. Use Series.plot() instead (GH18627)

  • Index.summary() is deprecated and will be removed in a future version (GH18217)

  • NDFrame.get_ftype_counts() is deprecated and will be removed in a future version (GH18243)

  • The convert_datetime64 parameter in DataFrame.to_records() has been deprecated and will be removed in a future version. The NumPy bug motivating this parameter has been resolved. The default value for this parameter has also changed from True to None (GH18160).

  • Series.rolling().apply(), DataFrame.rolling().apply(), Series.expanding().apply(), and DataFrame.expanding().apply() have deprecated passing an np.array by default. One will need to pass the new raw parameter to be explicit about what is passed (GH20584)

  • The data, base, strides, flags and itemsize properties of the Series and Index classes have been deprecated and will be removed in a future version (GH20419).

  • DatetimeIndex.offset is deprecated. Use DatetimeIndex.freq instead (GH20716)

  • Floor division between an integer ndarray and a Timedelta is deprecated. Divide by Timedelta.value instead (GH19761)

  • Setting PeriodIndex.freq (which was not guaranteed to work correctly) is deprecated. Use PeriodIndex.asfreq() instead (GH20678)

  • Index.get_duplicates() is deprecated and will be removed in a future version (GH20239)

  • The previous default behavior of negative indices in Categorical.take is deprecated. In a future version it will change from meaning missing values to meaning positional indices from the right. The future behavior is consistent with Series.take() (GH20664).

  • Passing multiple axes to the axis parameter in DataFrame.dropna() has been deprecated and will be removed in a future version (GH20987)

Removal of prior version deprecations/changes

  • Warnings against the obsolete usage Categorical(codes, categories), which were emitted for instance when the first two arguments to Categorical() had different dtypes, and recommended the use of Categorical.from_codes, have now been removed (GH8074)

  • The levels and labels attributes of a MultiIndex can no longer be set directly (GH4039).

  • pd.tseries.util.pivot_annual has been removed (deprecated since v0.19). Use pivot_table instead (GH18370)

  • pd.tseries.util.isleapyear has been removed (deprecated since v0.19). Use .is_leap_year property in Datetime-likes instead (GH18370)

  • pd.ordered_merge has been removed (deprecated since v0.19). Use pd.merge_ordered instead (GH18459)

  • The SparseList class has been removed (GH14007)

  • The pandas.io.wb and pandas.io.data stub modules have been removed (GH13735)

  • Categorical.from_array has been removed (GH13854)

  • The freq and how parameters have been removed from the rolling/expanding/ewm methods of DataFrame and Series (deprecated since v0.18). Instead, resample before calling the methods. (GH18601 & GH18668)

  • DatetimeIndex.to_datetime, Timestamp.to_datetime, PeriodIndex.to_datetime, and Index.to_datetime have been removed (GH8254, GH14096, GH14113)

  • read_csv() has dropped the skip_footer parameter (GH13386)

  • read_csv() has dropped the as_recarray parameter (GH13373)

  • read_csv() has dropped the buffer_lines parameter (GH13360)

  • read_csv() has dropped the compact_ints and use_unsigned parameters (GH13323)

  • The Timestamp class has dropped the offset attribute in favor of freq (GH13593)

  • The Series, Categorical, and Index classes have dropped the reshape method (GH13012)

  • pandas.tseries.frequencies.get_standard_freq has been removed in favor of pandas.tseries.frequencies.to_offset(freq).rule_code (GH13874)

  • The freqstr keyword has been removed from pandas.tseries.frequencies.to_offset in favor of freq (GH13874)

  • The Panel4D and PanelND classes have been removed (GH13776)

  • The Panel class has dropped the to_long and toLong methods (GH19077)

  • The options display.line_with and display.height are removed in favor of display.width and display.max_rows respectively (GH4391, GH19107)

  • The labels attribute of the Categorical class has been removed in favor of Categorical.codes (GH7768)

  • The flavor parameter have been removed from func:to_sql method (GH13611)

  • The modules pandas.tools.hashing and pandas.util.hashing have been removed (GH16223)

  • The top-level functions pd.rolling_*, pd.expanding_* and pd.ewm* have been removed (Deprecated since v0.18). Instead, use the DataFrame/Series methods rolling, expanding and ewm (GH18723)

  • Imports from pandas.core.common for functions such as is_datetime64_dtype are now removed. These are located in pandas.api.types. (GH13634, GH19769)

  • The infer_dst keyword in Series.tz_localize(), DatetimeIndex.tz_localize() and DatetimeIndex have been removed. infer_dst=True is equivalent to ambiguous='infer', and infer_dst=False to ambiguous='raise' (GH7963).

  • When .resample() was changed from an eager to a lazy operation, like .groupby() in v0.18.0, we put in place compatibility (with a FutureWarning), so operations would continue to work. This is now fully removed, so a Resampler will no longer forward compat operations (GH20554)

  • Remove long deprecated axis=None parameter from .replace() (GH20271)

Performance improvements

Documentation changes

Thanks to all of the contributors who participated in the pandas Documentation Sprint, which took place on March 10th. We had about 500 participants from over 30 locations across the world. You should notice that many of the API docstrings have greatly improved.

There were too many simultaneous contributions to include a release note for each improvement, but this GitHub search should give you an idea of how many docstrings were improved.

Special thanks to Marc Garcia for organizing the sprint. For more information, read the NumFOCUS blogpost recapping the sprint.

  • Changed spelling of “numpy” to “NumPy”, and “python” to “Python”. (GH19017)

  • Consistency when introducing code samples, using either colon or period. Rewrote some sentences for greater clarity, added more dynamic references to functions, methods and classes. (GH18941, GH18948, GH18973, GH19017)

  • Added a reference to DataFrame.assign() in the concatenate section of the merging documentation (GH18665)

Bug fixes



A class of bugs were introduced in pandas 0.21 with CategoricalDtype that affects the correctness of operations like merge, concat, and indexing when comparing multiple unordered Categorical arrays that have the same categories, but in a different order. We highly recommend upgrading or manually aligning your categories before doing these operations.

  • Bug in Categorical.equals returning the wrong result when comparing two unordered Categorical arrays with the same categories, but in a different order (GH16603)

  • Bug in pandas.api.types.union_categoricals() returning the wrong result when for unordered categoricals with the categories in a different order. This affected pandas.concat() with Categorical data (GH19096).

  • Bug in pandas.merge() returning the wrong result when joining on an unordered Categorical that had the same categories but in a different order (GH19551)

  • Bug in CategoricalIndex.get_indexer() returning the wrong result when target was an unordered Categorical that had the same categories as self but in a different order (GH19551)

  • Bug in Index.astype() with a categorical dtype where the resultant index is not converted to a CategoricalIndex for all types of index (GH18630)

  • Bug in Series.astype() and Categorical.astype() where an existing categorical data does not get updated (GH10696, GH18593)

  • Bug in Series.str.split() with expand=True incorrectly raising an IndexError on empty strings (GH20002).

  • Bug in Index constructor with dtype=CategoricalDtype(...) where categories and ordered are not maintained (GH19032)

  • Bug in Series constructor with scalar and dtype=CategoricalDtype(...) where categories and ordered are not maintained (GH19565)

  • Bug in Categorical.__iter__ not converting to Python types (GH19909)

  • Bug in pandas.factorize() returning the unique codes for the uniques. This now returns a Categorical with the same dtype as the input (GH19721)

  • Bug in pandas.factorize() including an item for missing values in the uniques return value (GH19721)

  • Bug in Series.take() with categorical data interpreting -1 in indices as missing value markers, rather than the last element of the Series (GH20664)


  • Bug in Series.__sub__() subtracting a non-nanosecond np.datetime64 object from a Series gave incorrect results (GH7996)

  • Bug in DatetimeIndex, TimedeltaIndex addition and subtraction of zero-dimensional integer arrays gave incorrect results (GH19012)

  • Bug in DatetimeIndex and TimedeltaIndex where adding or subtracting an array-like of DateOffset objects either raised (np.array, pd.Index) or broadcast incorrectly (pd.Series) (GH18849)

  • Bug in Series.__add__() adding Series with dtype timedelta64[ns] to a timezone-aware DatetimeIndex incorrectly dropped timezone information (GH13905)

  • Adding a Period object to a datetime or Timestamp object will now correctly raise a TypeError (GH17983)

  • Bug in Timestamp where comparison with an array of Timestamp objects would result in a RecursionError (GH15183)

  • Bug in Series floor-division where operating on a scalar timedelta raises an exception (GH18846)

  • Bug in DatetimeIndex where the repr was not showing high-precision time values at the end of a day (e.g., 23:59:59.999999999) (GH19030)

  • Bug in .astype() to non-ns timedelta units would hold the incorrect dtype (GH19176, GH19223, GH12425)

  • Bug in subtracting Series from NaT incorrectly returning NaT (GH19158)

  • Bug in Series.truncate() which raises TypeError with a monotonic PeriodIndex (GH17717)

  • Bug in pct_change() using periods and freq returned different length outputs (GH7292)

  • Bug in comparison of DatetimeIndex against None or datetime.date objects raising TypeError for == and != comparisons instead of all-False and all-True, respectively (GH19301)

  • Bug in Timestamp and to_datetime() where a string representing a barely out-of-bounds timestamp would be incorrectly rounded down instead of raising OutOfBoundsDatetime (GH19382)

  • Bug in Timestamp.floor() DatetimeIndex.floor() where time stamps far in the future and past were not rounded correctly (GH19206)

  • Bug in to_datetime() where passing an out-of-bounds datetime with errors='coerce' and utc=True would raise OutOfBoundsDatetime instead of parsing to NaT (GH19612)

  • Bug in DatetimeIndex and TimedeltaIndex addition and subtraction where name of the returned object was not always set consistently. (GH19744)

  • Bug in DatetimeIndex and TimedeltaIndex addition and subtraction where operations with numpy arrays raised TypeError (GH19847)

  • Bug in DatetimeIndex and TimedeltaIndex where setting the freq attribute was not fully supported (GH20678)


  • Bug in Timedelta.__mul__() where multiplying by NaT returned NaT instead of raising a TypeError (GH19819)

  • Bug in Series with dtype='timedelta64[ns]' where addition or subtraction of TimedeltaIndex had results cast to dtype='int64' (GH17250)

  • Bug in Series with dtype='timedelta64[ns]' where addition or subtraction of TimedeltaIndex could return a Series with an incorrect name (GH19043)

  • Bug in Timedelta.__floordiv__() and Timedelta.__rfloordiv__() dividing by many incompatible numpy objects was incorrectly allowed (GH18846)

  • Bug where dividing a scalar timedelta-like object with TimedeltaIndex performed the reciprocal operation (GH19125)

  • Bug in TimedeltaIndex where division by a Series would return a TimedeltaIndex instead of a Series (GH19042)

  • Bug in Timedelta.__add__(), Timedelta.__sub__() where adding or subtracting a np.timedelta64 object would return another np.timedelta64 instead of a Timedelta (GH19738)

  • Bug in Timedelta.__floordiv__(), Timedelta.__rfloordiv__() where operating with a Tick object would raise a TypeError instead of returning a numeric value (GH19738)

  • Bug in Period.asfreq() where periods near datetime(1, 1, 1) could be converted incorrectly (GH19643, GH19834)

  • Bug in Timedelta.total_seconds() causing precision errors, for example Timedelta('30S').total_seconds()==30.000000000000004 (GH19458)

  • Bug in Timedelta.__rmod__() where operating with a numpy.timedelta64 returned a timedelta64 object instead of a Timedelta (GH19820)

  • Multiplication of TimedeltaIndex by TimedeltaIndex will now raise TypeError instead of raising ValueError in cases of length mismatch (GH19333)

  • Bug in indexing a TimedeltaIndex with a np.timedelta64 object which was raising a TypeError (GH20393)


  • Bug in creating a Series from an array that contains both tz-naive and tz-aware values will result in a Series whose dtype is tz-aware instead of object (GH16406)

  • Bug in comparison of timezone-aware DatetimeIndex against NaT incorrectly raising TypeError (GH19276)

  • Bug in DatetimeIndex.astype() when converting between timezone aware dtypes, and converting from timezone aware to naive (GH18951)

  • Bug in comparing DatetimeIndex, which failed to raise TypeError when attempting to compare timezone-aware and timezone-naive datetimelike objects (GH18162)

  • Bug in localization of a naive, datetime string in a Series constructor with a datetime64[ns, tz] dtype (GH174151)

  • Timestamp.replace() will now handle Daylight Savings transitions gracefully (GH18319)

  • Bug in tz-aware DatetimeIndex where addition/subtraction with a TimedeltaIndex or array with dtype='timedelta64[ns]' was incorrect (GH17558)

  • Bug in DatetimeIndex.insert() where inserting NaT into a timezone-aware index incorrectly raised (GH16357)

  • Bug in DataFrame constructor, where tz-aware Datetimeindex and a given column name will result in an empty DataFrame (GH19157)

  • Bug in Timestamp.tz_localize() where localizing a timestamp near the minimum or maximum valid values could overflow and return a timestamp with an incorrect nanosecond value (GH12677)

  • Bug when iterating over DatetimeIndex that was localized with fixed timezone offset that rounded nanosecond precision to microseconds (GH19603)

  • Bug in DataFrame.diff() that raised an IndexError with tz-aware values (GH18578)

  • Bug in melt() that converted tz-aware dtypes to tz-naive (GH15785)

  • Bug in Dataframe.count() that raised an ValueError, if Dataframe.dropna() was called for a single column with timezone-aware values. (GH13407)


  • Bug in WeekOfMonth and Week where addition and subtraction did not roll correctly (GH18510, GH18672, GH18864)

  • Bug in WeekOfMonth and LastWeekOfMonth where default keyword arguments for constructor raised ValueError (GH19142)

  • Bug in FY5253Quarter, LastWeekOfMonth where rollback and rollforward behavior was inconsistent with addition and subtraction behavior (GH18854)

  • Bug in FY5253 where datetime addition and subtraction incremented incorrectly for dates on the year-end but not normalized to midnight (GH18854)

  • Bug in FY5253 where date offsets could incorrectly raise an AssertionError in arithmetic operations (GH14774)


  • Bug in Series constructor with an int or float list where specifying dtype=str, dtype='str' or dtype='U' failed to convert the data elements to strings (GH16605)

  • Bug in Index multiplication and division methods where operating with a Series would return an Index object instead of a Series object (GH19042)

  • Bug in the DataFrame constructor in which data containing very large positive or very large negative numbers was causing OverflowError (GH18584)

  • Bug in Index constructor with dtype='uint64' where int-like floats were not coerced to UInt64Index (GH18400)

  • Bug in DataFrame flex arithmetic (e.g. df.add(other, fill_value=foo)) with a fill_value other than None failed to raise NotImplementedError in corner cases where either the frame or other has length zero (GH19522)

  • Multiplication and division of numeric-dtyped Index objects with timedelta-like scalars returns TimedeltaIndex instead of raising TypeError (GH19333)

  • Bug where NaN was returned instead of 0 by Series.pct_change() and DataFrame.pct_change() when fill_method is not None (GH19873)


  • Bug in Series.str.get() with a dictionary in the values and the index not in the keys, raising KeyError (GH20671)


  • Bug in Index construction from list of mixed type tuples (GH18505)

  • Bug in Index.drop() when passing a list of both tuples and non-tuples (GH18304)

  • Bug in DataFrame.drop(), Panel.drop(), Series.drop(), Index.drop() where no KeyError is raised when dropping a non-existent element from an axis that contains duplicates (GH19186)

  • Bug in indexing a datetimelike Index that raised ValueError instead of IndexError (GH18386).

  • Index.to_series() now accepts index and name kwargs (GH18699)

  • DatetimeIndex.to_series() now accepts index and name kwargs (GH18699)

  • Bug in indexing non-scalar value from Series having non-unique Index will return value flattened (GH17610)

  • Bug in indexing with iterator containing only missing keys, which raised no error (GH20748)

  • Fixed inconsistency in .ix between list and scalar keys when the index has integer dtype and does not include the desired keys (GH20753)

  • Bug in __setitem__ when indexing a DataFrame with a 2-d boolean ndarray (GH18582)

  • Bug in str.extractall when there were no matches empty Index was returned instead of appropriate MultiIndex (GH19034)

  • Bug in IntervalIndex where empty and purely NA data was constructed inconsistently depending on the construction method (GH18421)

  • Bug in IntervalIndex.symmetric_difference() where the symmetric difference with a non-IntervalIndex did not raise (GH18475)

  • Bug in IntervalIndex where set operations that returned an empty IntervalIndex had the wrong dtype (GH19101)

  • Bug in DataFrame.drop_duplicates() where no KeyError is raised when passing in columns that don’t exist on the DataFrame (GH19726)

  • Bug in Index subclasses constructors that ignore unexpected keyword arguments (GH19348)

  • Bug in Index.difference() when taking difference of an Index with itself (GH20040)

  • Bug in DataFrame.first_valid_index() and DataFrame.last_valid_index() in presence of entire rows of NaNs in the middle of values (GH20499).

  • Bug in IntervalIndex where some indexing operations were not supported for overlapping or non-monotonic uint64 data (GH20636)

  • Bug in Series.is_unique where extraneous output in stderr is shown if Series contains objects with __ne__ defined (GH20661)

  • Bug in .loc assignment with a single-element list-like incorrectly assigns as a list (GH19474)

  • Bug in partial string indexing on a Series/DataFrame with a monotonic decreasing DatetimeIndex (GH19362)

  • Bug in performing in-place operations on a DataFrame with a duplicate Index (GH17105)

  • Bug in IntervalIndex.get_loc() and IntervalIndex.get_indexer() when used with an IntervalIndex containing a single interval (GH17284, GH20921)

  • Bug in .loc with a uint64 indexer (GH20722)




  • Better error message when attempting to plot but matplotlib is not installed (GH19810).

  • DataFrame.plot() now raises a ValueError when the x or y argument is improperly formed (GH18671)

  • Bug in DataFrame.plot() when x and y arguments given as positions caused incorrect referenced columns for line, bar and area plots (GH20056)

  • Bug in formatting tick labels with datetime.time() and fractional seconds (GH18478).

  • Series.plot.kde() has exposed the args ind and bw_method in the docstring (GH18461). The argument ind may now also be an integer (number of sample points).

  • DataFrame.plot() now supports multiple columns to the y argument (GH19699)


  • Bug when grouping by a single column and aggregating with a class like list or tuple (GH18079)

  • Fixed regression in DataFrame.groupby() which would not emit an error when called with a tuple key not in the index (GH18798)

  • Bug in DataFrame.resample() which silently ignored unsupported (or mistyped) options for label, closed and convention (GH19303)

  • Bug in DataFrame.groupby() where tuples were interpreted as lists of keys rather than as keys (GH17979, GH18249)

  • Bug in DataFrame.groupby() where aggregation by first/last/min/max was causing timestamps to lose precision (GH19526)

  • Bug in DataFrame.transform() where particular aggregation functions were being incorrectly cast to match the dtype(s) of the grouped data (GH19200)

  • Bug in DataFrame.groupby() passing the on= kwarg, and subsequently using .apply() (GH17813)

  • Bug in DataFrame.resample().aggregate not raising a KeyError when aggregating a non-existent column (GH16766, GH19566)

  • Bug in DataFrameGroupBy.cumsum() and DataFrameGroupBy.cumprod() when skipna was passed (GH19806)

  • Bug in DataFrame.resample() that dropped timezone information (GH13238)

  • Bug in DataFrame.groupby() where transformations using np.all and np.any were raising a ValueError (GH20653)

  • Bug in DataFrame.resample() where ffill, bfill, pad, backfill, fillna, interpolate, and asfreq were ignoring loffset. (GH20744)

  • Bug in DataFrame.groupby() when applying a function that has mixed data types and the user supplied function can fail on the grouping column (GH20949)

  • Bug in DataFrameGroupBy.rolling().apply() where operations performed against the associated DataFrameGroupBy object could impact the inclusion of the grouped item(s) in the result (GH14013)


  • Bug in which creating a SparseDataFrame from a dense Series or an unsupported type raised an uncontrolled exception (GH19374)

  • Bug in SparseDataFrame.to_csv causing exception (GH19384)

  • Bug in SparseSeries.memory_usage which caused segfault by accessing non sparse elements (GH19368)

  • Bug in constructing a SparseArray: if data is a scalar and index is defined it will coerce to float64 regardless of scalar’s dtype. (GH19163)


  • Bug in DataFrame.merge() where referencing a CategoricalIndex by name, where the by kwarg would KeyError (GH20777)

  • Bug in DataFrame.stack() which fails trying to sort mixed type levels under Python 3 (GH18310)

  • Bug in DataFrame.unstack() which casts int to float if columns is a MultiIndex with unused levels (GH17845)

  • Bug in DataFrame.unstack() which raises an error if index is a MultiIndex with unused labels on the unstacked level (GH18562)

  • Fixed construction of a Series from a dict containing NaN as key (GH18480)

  • Fixed construction of a DataFrame from a dict containing NaN as key (GH18455)

  • Disabled construction of a Series where len(index) > len(data) = 1, which previously would broadcast the data item, and now raises a ValueError (GH18819)

  • Suppressed error in the construction of a DataFrame from a dict containing scalar values when the corresponding keys are not included in the passed index (GH18600)

  • Fixed (changed from object to float64) dtype of DataFrame initialized with axes, no data, and dtype=int (GH19646)

  • Bug in Series.rank() where Series containing NaT modifies the Series inplace (GH18521)

  • Bug in cut() which fails when using readonly arrays (GH18773)

  • Bug in DataFrame.pivot_table() which fails when the aggfunc arg is of type string. The behavior is now consistent with other methods like agg and apply (GH18713)

  • Bug in DataFrame.merge() in which merging using Index objects as vectors raised an Exception (GH19038)

  • Bug in DataFrame.stack(), DataFrame.unstack(), Series.unstack() which were not returning subclasses (GH15563)

  • Bug in timezone comparisons, manifesting as a conversion of the index to UTC in .concat() (GH18523)

  • Bug in concat() when concatenating sparse and dense series it returns only a SparseDataFrame. Should be a DataFrame. (GH18914, GH18686, and GH16874)

  • Improved error message for DataFrame.merge() when there is no common merge key (GH19427)

  • Bug in DataFrame.join() which does an outer instead of a left join when being called with multiple DataFrames and some have non-unique indices (GH19624)

  • Series.rename() now accepts axis as a kwarg (GH18589)

  • Bug in rename() where an Index of same-length tuples was converted to a MultiIndex (GH19497)

  • Comparisons between Series and Index would return a Series with an incorrect name, ignoring the Index’s name attribute (GH19582)

  • Bug in qcut() where datetime and timedelta data with NaT present raised a ValueError (GH19768)

  • Bug in DataFrame.iterrows(), which would infers strings not compliant to ISO8601 to datetimes (GH19671)

  • Bug in Series constructor with Categorical where a ValueError is not raised when an index of different length is given (GH19342)

  • Bug in DataFrame.astype() where column metadata is lost when converting to categorical or a dictionary of dtypes (GH19920)

  • Bug in cut() and qcut() where timezone information was dropped (GH19872)

  • Bug in Series constructor with a dtype=str, previously raised in some cases (GH19853)

  • Bug in get_dummies(), and select_dtypes(), where duplicate column names caused incorrect behavior (GH20848)

  • Bug in isna(), which cannot handle ambiguous typed lists (GH20675)

  • Bug in concat() which raises an error when concatenating TZ-aware dataframes and all-NaT dataframes (GH12396)

  • Bug in concat() which raises an error when concatenating empty TZ-aware series (GH18447)



A total of 328 people contributed patches to this release. People with a “+” by their names contributed a patch for the first time.

  • Aaron Critchley

  • AbdealiJK +

  • Adam Hooper +

  • Albert Villanova del Moral

  • Alejandro Giacometti +

  • Alejandro Hohmann +

  • Alex Rychyk

  • Alexander Buchkovsky

  • Alexander Lenail +

  • Alexander Michael Schade

  • Aly Sivji +

  • Andreas Költringer +

  • Andrew

  • Andrew Bui +

  • András Novoszáth +

  • Andy Craze +

  • Andy R. Terrel

  • Anh Le +

  • Anil Kumar Pallekonda +

  • Antoine Pitrou +

  • Antonio Linde +

  • Antonio Molina +

  • Antonio Quinonez +

  • Armin Varshokar +

  • Artem Bogachev +

  • Avi Sen +

  • Azeez Oluwafemi +

  • Ben Auffarth +

  • Bernhard Thiel +

  • Bhavesh Poddar +

  • BielStela +

  • Blair +

  • Bob Haffner

  • Brett Naul +

  • Brock Mendel

  • Bryce Guinta +

  • Carlos Eduardo Moreira dos Santos +

  • Carlos García Márquez +

  • Carol Willing

  • Cheuk Ting Ho +

  • Chitrank Dixit +

  • Chris

  • Chris Burr +

  • Chris Catalfo +

  • Chris Mazzullo

  • Christian Chwala +

  • Cihan Ceyhan +

  • Clemens Brunner

  • Colin +

  • Cornelius Riemenschneider

  • Crystal Gong +

  • DaanVanHauwermeiren

  • Dan Dixey +

  • Daniel Frank +

  • Daniel Garrido +

  • Daniel Sakuma +

  • DataOmbudsman +

  • Dave Hirschfeld

  • Dave Lewis +

  • David Adrián Cañones Castellano +

  • David Arcos +

  • David C Hall +

  • David Fischer

  • David Hoese +

  • David Lutz +

  • David Polo +

  • David Stansby

  • Dennis Kamau +

  • Dillon Niederhut

  • Dimitri +

  • Dr. Irv

  • Dror Atariah

  • Eric Chea +

  • Eric Kisslinger

  • Eric O. LEBIGOT (EOL) +

  • FAN-GOD +

  • Fabian Retkowski +

  • Fer Sar +

  • Gabriel de Maeztu +

  • Gianpaolo Macario +

  • Giftlin Rajaiah

  • Gilberto Olimpio +

  • Gina +

  • Gjelt +

  • Graham Inggs +

  • Grant Roch

  • Grant Smith +

  • Grzegorz Konefał +

  • Guilherme Beltramini

  • HagaiHargil +

  • Hamish Pitkeathly +

  • Hammad Mashkoor +

  • Hannah Ferchland +

  • Hans

  • Haochen Wu +

  • Hissashi Rocha +

  • Iain Barr +

  • Ibrahim Sharaf ElDen +

  • Ignasi Fosch +

  • Igor Conrado Alves de Lima +

  • Igor Shelvinskyi +

  • Imanflow +

  • Ingolf Becker

  • Israel Saeta Pérez

  • Iva Koevska +

  • Jakub Nowacki +

  • Jan F-F +

  • Jan Koch +

  • Jan Werkmann

  • Janelle Zoutkamp +

  • Jason Bandlow +

  • Jaume Bonet +

  • Jay Alammar +

  • Jeff Reback

  • JennaVergeynst

  • Jimmy Woo +

  • Jing Qiang Goh +

  • Joachim Wagner +

  • Joan Martin Miralles +

  • Joel Nothman

  • Joeun Park +

  • John Cant +

  • Johnny Metz +

  • Jon Mease

  • Jonas Schulze +

  • Jongwony +

  • Jordi Contestí +

  • Joris Van den Bossche

  • José F. R. Fonseca +

  • Jovixe +

  • Julio Martinez +

  • Jörg Döpfert

  • KOBAYASHI Ittoku +

  • Kate Surta +

  • Kenneth +

  • Kevin Kuhl

  • Kevin Sheppard

  • Krzysztof Chomski

  • Ksenia +

  • Ksenia Bobrova +

  • Kunal Gosar +

  • Kurtis Kerstein +

  • Kyle Barron +

  • Laksh Arora +

  • Laurens Geffert +

  • Leif Walsh

  • Liam Marshall +

  • Liam3851 +

  • Licht Takeuchi

  • Liudmila +

  • Ludovico Russo +

  • Mabel Villalba +

  • Manan Pal Singh +

  • Manraj Singh

  • Marc +

  • Marc Garcia

  • Marco Hemken +

  • Maria del Mar Bibiloni +

  • Mario Corchero +

  • Mark Woodbridge +

  • Martin Journois +

  • Mason Gallo +

  • Matias Heikkilä +

  • Matt Braymer-Hayes

  • Matt Kirk +

  • Matt Maybeno +

  • Matthew Kirk +

  • Matthew Rocklin +

  • Matthew Roeschke

  • Matthias Bussonnier +

  • Max Mikhaylov +

  • Maxim Veksler +

  • Maximilian Roos

  • Maximiliano Greco +

  • Michael Penkov

  • Michael Röttger +

  • Michael Selik +

  • Michael Waskom

  • Mie~~~

  • Mike Kutzma +

  • Ming Li +

  • Mitar +

  • Mitch Negus +

  • Montana Low +

  • Moritz Münst +

  • Mortada Mehyar

  • Myles Braithwaite +

  • Nate Yoder

  • Nicholas Ursa +

  • Nick Chmura

  • Nikos Karagiannakis +

  • Nipun Sadvilkar +

  • Nis Martensen +

  • Noah +

  • Noémi Éltető +

  • Olivier Bilodeau +

  • Ondrej Kokes +

  • Onno Eberhard +

  • Paul Ganssle +

  • Paul Mannino +

  • Paul Reidy

  • Paulo Roberto de Oliveira Castro +

  • Pepe Flores +

  • Peter Hoffmann

  • Phil Ngo +

  • Pietro Battiston

  • Pranav Suri +

  • Priyanka Ojha +

  • Pulkit Maloo +

  • README Bot +

  • Ray Bell +

  • Riccardo Magliocchetti +

  • Ridhwan Luthra +

  • Robert Meyer

  • Robin

  • Robin Kiplang’at +

  • Rohan Pandit +

  • Rok Mihevc +

  • Rouz Azari

  • Ryszard T. Kaleta +

  • Sam Cohan

  • Sam Foo

  • Samir Musali +

  • Samuel Sinayoko +

  • Sangwoong Yoon

  • SarahJessica +

  • Sharad Vijalapuram +

  • Shubham Chaudhary +

  • SiYoungOh +

  • Sietse Brouwer

  • Simone Basso +

  • Stefania Delprete +

  • Stefano Cianciulli +

  • Stephen Childs +

  • StephenVoland +

  • Stijn Van Hoey +

  • Sven

  • Talitha Pumar +

  • Tarbo Fukazawa +

  • Ted Petrou +

  • Thomas A Caswell

  • Tim Hoffmann +

  • Tim Swast

  • Tom Augspurger

  • Tommy +

  • Tulio Casagrande +

  • Tushar Gupta +

  • Tushar Mittal +

  • Upkar Lidder +

  • Victor Villas +

  • Vince W +

  • Vinícius Figueiredo +

  • Vipin Kumar +

  • WBare

  • Wenhuan +

  • Wes Turner

  • William Ayd

  • Wilson Lin +

  • Xbar

  • Yaroslav Halchenko

  • Yee Mey

  • Yeongseon Choe +

  • Yian +

  • Yimeng Zhang

  • ZhuBaohe +

  • Zihao Zhao +

  • adatasetaday +

  • akielbowicz +

  • akosel +

  • alinde1 +

  • amuta +

  • bolkedebruin

  • cbertinato

  • cgohlke

  • charlie0389 +

  • chris-b1

  • csfarkas +

  • dajcs +

  • deflatSOCO +

  • derestle-htwg

  • discort

  • dmanikowski-reef +

  • donK23 +

  • elrubio +

  • fivemok +

  • fjdiod

  • fjetter +

  • froessler +

  • gabrielclow

  • gfyoung

  • ghasemnaddaf

  • h-vetinari +

  • himanshu awasthi +

  • ignamv +

  • jayfoad +

  • jazzmuesli +

  • jbrockmendel

  • jen w +

  • jjames34 +

  • joaoavf +

  • joders +

  • jschendel

  • juan huguet +

  • l736x +

  • luzpaz +

  • mdeboc +

  • miguelmorin +

  • miker985

  • miquelcamprodon +

  • orereta +

  • ottiP +

  • peterpanmj +

  • rafarui +

  • raph-m +

  • readyready15728 +

  • rmihael +

  • samghelms +

  • scriptomation +

  • sfoo +

  • stefansimik +

  • stonebig

  • tmnhat2001 +

  • tomneep +

  • topper-123

  • tv3141 +

  • verakai +

  • xpvpc +

  • zhanghui +