ENH: Add option to use nullable dtypes in read_csv (#48776)

* ENH: Add option to use nullable dtypes in read_csv

* Finish implementation

* Update

* Fix mypy

* Add tests and fix call

* Fix typo

Author: phofl

doc/source/user_guide/io.rst

@@ -197,6 +197,14 @@ dtype : Type name or dict of column -> type, default ``None``
      Support for defaultdict was added. Specify a defaultdict as input where
      the default determines the dtype of the columns which are not explicitly
      listed.
+
+use_nullable_dtypes : bool = False
+    Whether or not to use nullable dtypes as default when reading data. If
+    set to True, nullable dtypes are used for all dtypes that have a nullable
+    implementation, even if no nulls are present.
+
+    .. versionadded:: 2.0
+
 engine : {``'c'``, ``'python'``, ``'pyarrow'``}
   Parser engine to use. The C and pyarrow engines are faster, while the python engine
   is currently more feature-complete. Multithreading is currently only supported by

doc/source/whatsnew/v1.6.0.rst

@@ -32,6 +32,7 @@ Other enhancements
 - :meth:`.DataFrameGroupBy.quantile` and :meth:`.SeriesGroupBy.quantile` now preserve nullable dtypes instead of casting to numpy dtypes (:issue:`37493`)
 - :meth:`Series.add_suffix`, :meth:`DataFrame.add_suffix`, :meth:`Series.add_prefix` and :meth:`DataFrame.add_prefix` support an ``axis`` argument. If ``axis`` is set, the default behaviour of which axis to consider can be overwritten (:issue:`47819`)
 - :func:`assert_frame_equal` now shows the first element where the DataFrames differ, analogously to ``pytest``'s output (:issue:`47910`)
+- Added new argument ``use_nullable_dtypes`` to :func:`read_csv` to enable automatic conversion to nullable dtypes (:issue:`36712`)
 - Added ``index`` parameter to :meth:`DataFrame.to_dict` (:issue:`46398`)
 - Added metadata propagation for binary operators on :class:`DataFrame` (:issue:`28283`)
 - :class:`.CategoricalConversionWarning`, :class:`.InvalidComparison`, :class:`.InvalidVersion`, :class:`.LossySetitemError`, and :class:`.NoBufferPresent` are now exposed in ``pandas.errors`` (:issue:`27656`)

pandas/_libs/parsers.pyx

@@ -342,6 +342,7 @@ cdef class TextReader:
         object index_col
         object skiprows
         object dtype
+        bint use_nullable_dtypes
         object usecols
         set unnamed_cols  # set[str]
 
@@ -380,7 +381,8 @@ cdef class TextReader:
                   bint mangle_dupe_cols=True,
                   float_precision=None,
                   bint skip_blank_lines=True,
-                  encoding_errors=b"strict"):
+                  encoding_errors=b"strict",
+                  use_nullable_dtypes=False):
 
         # set encoding for native Python and C library
         if isinstance(encoding_errors, str):
@@ -505,6 +507,7 @@ cdef class TextReader:
         # - DtypeObj
         # - dict[Any, DtypeObj]
         self.dtype = dtype
+        self.use_nullable_dtypes = use_nullable_dtypes
 
         # XXX
         self.noconvert = set()
@@ -933,6 +936,7 @@ cdef class TextReader:
             bint na_filter = 0
             int64_t num_cols
             dict result
+            bint use_nullable_dtypes
 
         start = self.parser_start
 
@@ -1053,8 +1057,14 @@ cdef class TextReader:
                     self._free_na_set(na_hashset)
 
             # don't try to upcast EAs
-            if na_count > 0 and not is_extension_array_dtype(col_dtype):
-                col_res = _maybe_upcast(col_res)
+            if (
+                na_count > 0 and not is_extension_array_dtype(col_dtype)
+                or self.use_nullable_dtypes
+            ):
+                use_nullable_dtypes = self.use_nullable_dtypes and col_dtype is None
+                col_res = _maybe_upcast(
+                    col_res, use_nullable_dtypes=use_nullable_dtypes
+                )
 
             if col_res is None:
                 raise ParserError(f'Unable to parse column {i}')

pandas/io/parsers/base_parser.py

@@ -15,6 +15,7 @@
     Hashable,
     Iterable,
     List,
+    Literal,
     Mapping,
     Sequence,
     Tuple,
@@ -50,6 +51,7 @@
     is_dict_like,
     is_dtype_equal,
     is_extension_array_dtype,
+    is_float_dtype,
     is_integer,
     is_integer_dtype,
     is_list_like,
@@ -61,8 +63,14 @@
 from pandas.core.dtypes.dtypes import CategoricalDtype
 from pandas.core.dtypes.missing import isna
 
+from pandas import StringDtype
 from pandas.core import algorithms
-from pandas.core.arrays import Categorical
+from pandas.core.arrays import (
+    BooleanArray,
+    Categorical,
+    FloatingArray,
+    IntegerArray,
+)
 from pandas.core.indexes.api import (
     Index,
     MultiIndex,
@@ -110,6 +118,7 @@ def __init__(self, kwds) -> None:
 
         self.dtype = copy(kwds.get("dtype", None))
         self.converters = kwds.get("converters")
+        self.use_nullable_dtypes = kwds.get("use_nullable_dtypes", False)
 
         self.true_values = kwds.get("true_values")
         self.false_values = kwds.get("false_values")
@@ -508,7 +517,7 @@ def _agg_index(self, index, try_parse_dates: bool = True) -> Index:
             )
 
             arr, _ = self._infer_types(
-                arr, col_na_values | col_na_fvalues, try_num_bool
+                arr, col_na_values | col_na_fvalues, cast_type is None, try_num_bool
             )
             arrays.append(arr)
 
@@ -574,7 +583,10 @@ def _convert_to_ndarrays(
                     values = lib.map_infer_mask(values, conv_f, mask)
 
                 cvals, na_count = self._infer_types(
-                    values, set(col_na_values) | col_na_fvalues, try_num_bool=False
+                    values,
+                    set(col_na_values) | col_na_fvalues,
+                    cast_type is None,
+                    try_num_bool=False,
                 )
             else:
                 is_ea = is_extension_array_dtype(cast_type)
@@ -585,14 +597,14 @@ def _convert_to_ndarrays(
 
                 # general type inference and conversion
                 cvals, na_count = self._infer_types(
-                    values, set(col_na_values) | col_na_fvalues, try_num_bool
+                    values,
+                    set(col_na_values) | col_na_fvalues,
+                    cast_type is None,
+                    try_num_bool,
                 )
 
                 # type specified in dtype param or cast_type is an EA
-                if cast_type and (
-                    not is_dtype_equal(cvals, cast_type)
-                    or is_extension_array_dtype(cast_type)
-                ):
+                if cast_type and (not is_dtype_equal(cvals, cast_type) or is_ea):
                     if not is_ea and na_count > 0:
                         try:
                             if is_bool_dtype(cast_type):
@@ -679,14 +691,17 @@ def _set(x) -> int:
 
         return noconvert_columns
 
-    def _infer_types(self, values, na_values, try_num_bool: bool = True):
+    def _infer_types(
+        self, values, na_values, no_dtype_specified, try_num_bool: bool = True
+    ):
         """
         Infer types of values, possibly casting
 
         Parameters
         ----------
         values : ndarray
         na_values : set
+        no_dtype_specified: Specifies if we want to cast explicitly
         try_num_bool : bool, default try
            try to cast values to numeric (first preference) or boolean
 
@@ -707,28 +722,62 @@ def _infer_types(self, values, na_values, try_num_bool: bool = True):
                 np.putmask(values, mask, np.nan)
             return values, na_count
 
+        use_nullable_dtypes: Literal[True] | Literal[False] = (
+            self.use_nullable_dtypes and no_dtype_specified
+        )
+        result: ArrayLike
+
         if try_num_bool and is_object_dtype(values.dtype):
             # exclude e.g DatetimeIndex here
             try:
-                result, _ = lib.maybe_convert_numeric(values, na_values, False)
+                result, result_mask = lib.maybe_convert_numeric(
+                    values,
+                    na_values,
+                    False,
+                    convert_to_masked_nullable=use_nullable_dtypes,
+                )
             except (ValueError, TypeError):
                 # e.g. encountering datetime string gets ValueError
                 #  TypeError can be raised in floatify
+                na_count = parsers.sanitize_objects(values, na_values)
                 result = values
-                na_count = parsers.sanitize_objects(result, na_values)
             else:
-                na_count = isna(result).sum()
+                if use_nullable_dtypes:
+                    if result_mask is None:
+                        result_mask = np.zeros(result.shape, dtype=np.bool_)
+
+                    if result_mask.all():
+                        result = IntegerArray(
+                            np.ones(result_mask.shape, dtype=np.int64), result_mask
+                        )
+                    elif is_integer_dtype(result):
+                        result = IntegerArray(result, result_mask)
+                    elif is_bool_dtype(result):
+                        result = BooleanArray(result, result_mask)
+                    elif is_float_dtype(result):
+                        result = FloatingArray(result, result_mask)
+
+                    na_count = result_mask.sum()
+                else:
+                    na_count = isna(result).sum()
         else:
             result = values
             if values.dtype == np.object_:
                 na_count = parsers.sanitize_objects(values, na_values)
 
         if result.dtype == np.object_ and try_num_bool:
-            result, _ = libops.maybe_convert_bool(
+            result, bool_mask = libops.maybe_convert_bool(
                 np.asarray(values),
                 true_values=self.true_values,
                 false_values=self.false_values,
+                convert_to_masked_nullable=use_nullable_dtypes,
             )
+            if result.dtype == np.bool_ and use_nullable_dtypes:
+                if bool_mask is None:
+                    bool_mask = np.zeros(result.shape, dtype=np.bool_)
+                result = BooleanArray(result, bool_mask)
+            elif result.dtype == np.object_ and use_nullable_dtypes:
+                result = StringDtype().construct_array_type()._from_sequence(values)
 
         return result, na_count
 
@@ -1146,6 +1195,7 @@ def converter(*date_cols):
     "on_bad_lines": ParserBase.BadLineHandleMethod.ERROR,
     "error_bad_lines": None,
     "warn_bad_lines": None,
+    "use_nullable_dtypes": False,
 }
 
 

pandas/io/parsers/readers.py

@@ -427,6 +427,13 @@
 
     .. versionadded:: 1.2
 
+use_nullable_dtypes : bool = False
+    Whether or not to use nullable dtypes as default when reading data. If
+    set to True, nullable dtypes are used for all dtypes that have a nullable
+    implementation, even if no nulls are present.
+
+    .. versionadded:: 2.0
+
 Returns
 -------
 DataFrame or TextFileReader
@@ -669,6 +676,7 @@ def read_csv(
     memory_map: bool = ...,
     float_precision: Literal["high", "legacy"] | None = ...,
     storage_options: StorageOptions = ...,
+    use_nullable_dtypes: bool = ...,
 ) -> TextFileReader:
     ...
 
@@ -729,6 +737,7 @@ def read_csv(
     memory_map: bool = ...,
     float_precision: Literal["high", "legacy"] | None = ...,
     storage_options: StorageOptions = ...,
+    use_nullable_dtypes: bool = ...,
 ) -> TextFileReader:
     ...
 
@@ -789,6 +798,7 @@ def read_csv(
     memory_map: bool = ...,
     float_precision: Literal["high", "legacy"] | None = ...,
     storage_options: StorageOptions = ...,
+    use_nullable_dtypes: bool = ...,
 ) -> DataFrame:
     ...
 
@@ -849,6 +859,7 @@ def read_csv(
     memory_map: bool = ...,
     float_precision: Literal["high", "legacy"] | None = ...,
     storage_options: StorageOptions = ...,
+    use_nullable_dtypes: bool = ...,
 ) -> DataFrame | TextFileReader:
     ...
 
@@ -928,6 +939,7 @@ def read_csv(
     memory_map: bool = False,
     float_precision: Literal["high", "legacy"] | None = None,
     storage_options: StorageOptions = None,
+    use_nullable_dtypes: bool = False,
 ) -> DataFrame | TextFileReader:
     # locals() should never be modified
     kwds = locals().copy()
@@ -1008,6 +1020,7 @@ def read_table(
     memory_map: bool = ...,
     float_precision: str | None = ...,
     storage_options: StorageOptions = ...,
+    use_nullable_dtypes: bool = ...,
 ) -> TextFileReader:
     ...
 
@@ -1068,6 +1081,7 @@ def read_table(
     memory_map: bool = ...,
     float_precision: str | None = ...,
     storage_options: StorageOptions = ...,
+    use_nullable_dtypes: bool = ...,
 ) -> TextFileReader:
     ...
 
@@ -1128,6 +1142,7 @@ def read_table(
     memory_map: bool = ...,
     float_precision: str | None = ...,
     storage_options: StorageOptions = ...,
+    use_nullable_dtypes: bool = ...,
 ) -> DataFrame:
     ...
 
@@ -1188,6 +1203,7 @@ def read_table(
     memory_map: bool = ...,
     float_precision: str | None = ...,
     storage_options: StorageOptions = ...,
+    use_nullable_dtypes: bool = ...,
 ) -> DataFrame | TextFileReader:
     ...
 
@@ -1267,6 +1283,7 @@ def read_table(
     memory_map: bool = False,
     float_precision: str | None = None,
     storage_options: StorageOptions = None,
+    use_nullable_dtypes: bool = False,
 ) -> DataFrame | TextFileReader:
     # locals() should never be modified
     kwds = locals().copy()