Series round trip to pandas with arrow types copies data

[datapythonista]

Polars version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of Polars.

Issue description

I'd expect that a polars Series doesn't copy data in a roundtrip to pandas when arrow types are used:

polars.from_pandas(data.to_pandas(use_pyarrow_extension_array=True))

Interesting enough this is true when numpy types are used with use_pyarrow_extension_array=False, but not when data is always an arrow array.

I see in @ghuls comment that this is expected to be true: #6756 (comment)

But as shown in the reproducible example it seems to be making a copy.

Reproducible example

original_polars_series = polars.Series([0, 1, 2, 3, 4, 5, 6, 7, 8, 9] * 10_000_000)
print('original polars, address:', original_polars_series._get_ptr())

pandas_with_arrow = original_polars_series.to_pandas(use_pyarrow_extension_array=True)
print('pandas exported with arrow types, address:', pandas_with_arrow._data.array._data.chunks[0].buffers()[1].address)

polars_roundtrip_with_arrow = polars.from_pandas(pandas_with_arrow)
print('polars after roundtrip to pandas with arrow types, address:', polars_roundtrip_with_arrow._get_ptr())

pandas_with_numpy = original_polars_series.to_pandas(use_pyarrow_extension_array=False)
print('pandas exported with numpy types, address:', pandas_with_numpy.to_numpy().__array_interface__['data'][0])

polars_roundtrip_with_numpy = polars.from_pandas(pandas_with_numpy)
print('polars after roundtrip to pandas with numpy types, address:', polars_roundtrip_with_numpy._get_ptr())

Expected behavior

original polars, address: 140026900974656
pandas exported with arrow types, address: 140026900974656
polars after roundtrip to pandas with arrow types, address: 140028774713088     <--- I'd expect this address to stay the same as the original
pandas exported with numpy types, address: 140026900974656
polars after roundtrip to pandas with numpy types, address: 140026900974656

Installed versions

---Version info---
Polars: 0.16.6
Index type: UInt32
Platform: Linux-6.1.10-arch1-1-x86_64-with-glibc2.10
Python: 3.8.15 | packaged by conda-forge | (default, Nov 22 2022, 08:49:35) 
[GCC 10.4.0]
---Optional dependencies---
pyarrow: 10.0.1
pandas: 2.0.0.dev0+1580.g686d674f2d
numpy: 1.23.5
fsspec: 2022.11.0
connectorx: <not installed>
xlsx2csv: <not installed>
deltalake: <not installed>
matplotlib: 3.6.3```

</details>

[datapythonista]

Timings also seem to confirm the copy with arrow types and not with numpy:

>>> %timeit _ = polars.from_pandas(original_polars_series.to_pandas(use_pyarrow_extension_array=False))
393 µs ± 3.9 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)
>>> %timeit _ = polars.from_pandas(original_polars_series.to_pandas(use_pyarrow_extension_array=True))
141 ms ± 3.72 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

[ghuls]

@datapythonista I assume this is a pyarrow limitation at the moment.

# Converting Pandas Series with PyArrow results in a copy.
In [46]: pl.from_arrow(pa.array(pandas_with_arrow))._get_ptr()
Out[46]: 140342545672704

In [47]: pl.from_arrow(pa.array(pandas_with_arrow))._get_ptr()
Out[47]: 140342545671424

In [48]: pl.from_arrow(pa.array(pandas_with_arrow))._get_ptr()
Out[48]: 140342545670976

# Converting PyArrow extension array backing Pandas Series avoid the copy.
In [78]: pl.from_arrow(pandas_with_arrow._data.arrays[0]._data.chunks[0])._get_ptr()
Out[78]: 140347645945408

In [79]: pl.from_arrow(pandas_with_arrow._data.arrays[0]._data.chunks[0])._get_ptr()
Out[79]: 140347645945408

In [80]: pl.from_arrow(pandas_with_arrow._data.arrays[0]._data.chunks[0])._get_ptr()
Out[80]: 140347645945408

In [81]: pl.from_arrow(pandas_with_arrow._data.arrays[0]._data.chunks[0])
Out[81]:
shape: (100000000,)
Series: '' [i64]
[
        0
        1
        2
        3
        4
        5
        6
        7
        8
        9
        0
        1
        ...
        7
        8
        9
        0
        1
        2
        3
        4
        5
        6
        7
        8
        9
]

[datapythonista]

Seems like the problem is in Polars, which is making the copy by combining the arrow chunks even when there is just one. I opened #7084 to fix it.