feat(python): Implement DataFrame Interchange Protocol through pyarrow

py-polars/docs/source/reference/dataframe/export.rst

@@ -8,6 +8,7 @@ Export DataFrame data to other formats:
 .. autosummary::
    :toctree: api/
 
+    DataFrame.__dataframe__
     DataFrame.to_arrow
     DataFrame.to_dict
     DataFrame.to_dicts

py-polars/docs/source/reference/functions.rst

@@ -9,6 +9,7 @@ Conversion
    :toctree: api/
 
     from_arrow
+    from_dataframe
     from_dict
     from_dicts
     from_numpy

py-polars/polars/__init__.py

@@ -16,6 +16,7 @@ def version() -> str:
 from polars.cfg import Config
 from polars.convert import (
     from_arrow,
+    from_dataframe,
     from_dict,
     from_dicts,
     from_numpy,
@@ -291,6 +292,7 @@ def version() -> str:
     "duration",
     "coalesce",
     # polars.convert
+    "from_dataframe",
     "from_dict",
     "from_dicts",
     "from_records",

py-polars/polars/convert.py

@@ -3,6 +3,7 @@
 from typing import TYPE_CHECKING, Any, Mapping, Sequence, overload
 
 from polars.datatypes import N_INFER_DEFAULT, SchemaDefinition, SchemaDict
+from polars.dependencies import _PYARROW_AVAILABLE
 from polars.dependencies import numpy as np
 from polars.dependencies import pandas as pd
 from polars.dependencies import pyarrow as pa
@@ -479,3 +480,50 @@ def from_pandas(
         )
     else:
         raise ValueError(f"Expected pandas DataFrame or Series, got {type(df)}.")
+
+
+def from_dataframe(df: Any, allow_copy: bool = True) -> DataFrame:
+    """
+    Build a Polars DataFrame from any dataframe supporting the interchange protocol.
+
+    Parameters
+    ----------
+    df
+        Object supporting the dataframe interchange protocol, i.e. must have implemented
+        the ``__dataframe__`` method.
+    allow_copy
+        Allow memory to be copied to perform the conversion. If set to False, causes
+        conversions that are not zero-copy to fail.
+
+    Notes
+    -----
+    Details on the dataframe interchange protocol:
+    https://data-apis.org/dataframe-protocol/latest/index.html
+
+    Zero-copy conversions currently cannot be guaranteed and will throw a
+    ``NotImplementedError``.
+
+    Using a dedicated function like :func:`from_pandas` or :func:`from_arrow` is a more
+    efficient method of conversion.
+
+    """
+    if isinstance(df, DataFrame):
+        return df
+    if not hasattr(df, "__dataframe__"):
+        raise TypeError(
+            f"`df` of type {type(df)} does not support the dataframe interchange"
+            " protocol."
+        )
+    if not _PYARROW_AVAILABLE or int(pa.__version__.split(".")[0]) < 11:
+        raise ImportError(
+            "pyarrow>=11.0.0 is required for converting a dataframe interchange object"
+            " to a Polars dataframe."
+        )
+    if not allow_copy:
+        raise NotImplementedError(
+            "Polars cannot guarantee zero-copy conversion from dataframe interchange"
+            " objects at this time."
+        )
+
+    pa_table = pa.interchange.from_dataframe(df, allow_copy=allow_copy)
+    return from_arrow(pa_table, rechunk=allow_copy)  # type: ignore[return-value]

py-polars/polars/internals/dataframe/frame.py

@@ -33,6 +33,7 @@
     INTEGER_DTYPES,
     N_INFER_DEFAULT,
     Boolean,
+    Categorical,
     DataTypeClass,
     Float64,
     Int8,
@@ -105,6 +106,8 @@
     from typing_extensions import Concatenate, ParamSpec, TypeAlias
 
 if TYPE_CHECKING:
+    from pyarrow.interchange.dataframe import _PyArrowDataFrame
+
     from polars.internals.type_aliases import (
         AsofJoinStrategy,
         AvroCompression,
@@ -1126,6 +1129,42 @@ def schema(self) -> SchemaDict:
         """
         return dict(zip(self.columns, self.dtypes))
 
+    def __dataframe__(
+        self, nan_as_null: bool = False, allow_copy: bool = True
+    ) -> _PyArrowDataFrame:
+        """
+        Convert to a dataframe object implementing the dataframe interchange protocol.
+
+        Parameters
+        ----------
+        nan_as_null
+            Overwrite null values in the data with ``NaN``.
+        allow_copy
+            Allow memory to be copied to perform the conversion. If set to False, causes
+            conversions that are not zero-copy to fail.
+
+        Notes
+        -----
+        Details on the dataframe interchange protocol:
+        https://data-apis.org/dataframe-protocol/latest/index.html
+
+        `nan_as_null` currently has no effect; once support for nullable extension
+        dtypes is added, this value should be propagated to columns.
+
+        """
+        if not _PYARROW_AVAILABLE or int(pa.__version__.split(".")[0]) < 11:
+            raise ImportError(
+                "pyarrow>=11.0.0 is required for converting a Polars dataframe to a"
+                " dataframe interchange object."
+            )
+        if not allow_copy and Categorical in self.schema.values():
+            raise NotImplementedError(
+                "Polars does not offer zero-copy conversion to Arrow for categorical"
+                " columns. Set `allow_copy=True` or cast categorical columns to"
+                " string first."
+            )
+        return self.to_arrow().__dataframe__(nan_as_null, allow_copy)
+
     def _comp(self, other: Any, op: ComparisonOperator) -> DataFrame:
         """Compare a DataFrame with another object."""
         if isinstance(other, DataFrame):

py-polars/tests/unit/test_interchange.py

@@ -0,0 +1,118 @@
+import sys
+
+import pandas as pd
+import pytest
+from _pytest.monkeypatch import MonkeyPatch
+
+import polars as pl
+from polars.testing import assert_frame_equal
+
+
+def test_interchange() -> None:
+    df = pl.DataFrame({"a": [1, 2], "b": [3.0, 4.0], "c": ["foo", "bar"]})
+    dfi = df.__dataframe__()
+
+    # Testing some random properties to make sure conversion happened correctly
+    assert dfi.num_rows() == 2
+    assert dfi.get_column(0).dtype[1] == 64
+    assert dfi.get_column_by_name("c").get_buffers()["data"][0].bufsize == 6
+
+
+def test_interchange_pyarrow_required(monkeypatch: MonkeyPatch) -> None:
+    monkeypatch.setattr(pl.internals.dataframe.frame, "_PYARROW_AVAILABLE", False)
+
+    df = pl.DataFrame({"a": [1, 2]})
+    with pytest.raises(ImportError, match="pyarrow"):
+        df.__dataframe__()
+
+
+def test_interchange_pyarrow_min_version(monkeypatch: MonkeyPatch) -> None:
+    monkeypatch.setattr(
+        pl.internals.dataframe.frame.pa,  # type: ignore[attr-defined]
+        "__version__",
+        "10.0.0",
+    )
+
+    df = pl.DataFrame({"a": [1, 2]})
+    with pytest.raises(ImportError, match="pyarrow"):
+        df.__dataframe__()
+
+
+def test_interchange_categorical() -> None:
+    df = pl.DataFrame({"a": ["foo", "bar"]}, schema={"a": pl.Categorical})
+
+    # Conversion requires copy
+    dfi = df.__dataframe__(allow_copy=True)
+    assert dfi.get_column_by_name("a").dtype[0] == 23  # 23 signifies categorical dtype
+
+    # If copy not allowed, throws an error
+    with pytest.raises(NotImplementedError, match="categorical"):
+        df.__dataframe__(allow_copy=False)
+
+
+def test_from_dataframe() -> None:
+    df = pl.DataFrame({"a": [1, 2], "b": [3.0, 4.0], "c": ["foo", "bar"]})
+    dfi = df.__dataframe__()
+    result = pl.from_dataframe(dfi)
+    assert_frame_equal(result, df)
+
+
+@pytest.mark.xfail(
+    sys.version_info < (3, 8),
+    reason="Pandas does not implement the protocol on Python 3.7",
+)
+def test_from_dataframe_pandas() -> None:
+    data = {"a": [1, 2], "b": [3.0, 4.0], "c": ["foo", "bar"]}
+
+    # Pandas dataframe
+    df = pd.DataFrame(data)
+    result = pl.from_dataframe(df)
+    expected = pl.DataFrame(data)
+    assert_frame_equal(result, expected)
+
+
+@pytest.mark.xfail(
+    sys.version_info < (3, 8),
+    reason="Pandas does not implement the protocol on Python 3.7",
+)
+def test_from_dataframe_allow_copy() -> None:
+    # Zero copy only allowed when input is already a Polars dataframe
+    df = pl.DataFrame({"a": [1, 2]})
+    result = pl.from_dataframe(df, allow_copy=True)
+    assert_frame_equal(result, df)
+
+    # Zero copy cannot be guaranteed for other inputs at this time
+    df_pandas = pd.DataFrame({"a": [1, 2]})
+    with pytest.raises(NotImplementedError):
+        pl.from_dataframe(df_pandas, allow_copy=False)
+
+
+def test_from_dataframe_invalid_type() -> None:
+    df = [[1, 2], [3, 4]]
+    with pytest.raises(TypeError):
+        pl.from_dataframe(df)
+
+
+def test_from_dataframe_pyarrow_required(monkeypatch: MonkeyPatch) -> None:
+    monkeypatch.setattr(pl.convert, "_PYARROW_AVAILABLE", False)
+
+    df = pl.DataFrame({"a": [1, 2]})
+    with pytest.raises(ImportError, match="pyarrow"):
+        pl.from_dataframe(df.__dataframe__())
+
+    # 'Converting' from a Polars dataframe does not hit this requirement
+    result = pl.from_dataframe(df)
+    assert_frame_equal(result, df)
+
+
+def test_from_dataframe_pyarrow_min_version(monkeypatch: MonkeyPatch) -> None:
+    dfi = pl.DataFrame({"a": [1, 2]}).__dataframe__()
+
+    monkeypatch.setattr(
+        pl.convert.pa,  # type: ignore[attr-defined]
+        "__version__",
+        "10.0.0",
+    )
+
+    with pytest.raises(ImportError, match="pyarrow"):
+        pl.from_dataframe(dfi)