[nleerollergh]

Rationale for this change

Five Parquet metadata/schema equals() wrappers (Statistics, ColumnChunkMetaData, RowGroupMetaData, ParquetSchema, ColumnSchema) accept None for their typed other argument and then dereference a native pointer on it, which can crash the interpreter. This follows the existing FileMetaData.equals() pattern, which already declares other not None.

What changes are included in this PR?

• Require non-null typed arguments in the five affected Cython equals() methods. __eq__ already catches the resulting TypeError and returns NotImplemented, so obj == None is False and obj != None is True.
• Add a regression test covering ==, != and direct .equals(None) for all five wrappers.

Are these changes tested?

Yes. I added a regression test and ran it locally: it passes with the fix and segfaults without it (details in the comment below). flake8 and cython-lint pass on the changed files.

See comment for detailed testing evidence.

Are there any user-facing changes?

obj.equals(None) now raises TypeError instead of risking a crash. obj == None returns False.

This PR contains a "Critical Fix". It fixes a bug that causes a crash (category c): calling equals(None), or comparing with ==/!= against None, on Statistics, ColumnChunkMetaData, RowGroupMetaData, ParquetSchema or ColumnSchema could dereference a null native pointer and crash the Python interpreter instead of raising an error.

Was AI used for this PR?

Yes, I wrote the initial implementation (the not None change to the five equals() methods, following the existing FileMetaData.equals() pattern) and a first version of the regression test. I then used AI (Claude Code) to review and refine it before then structuring the commit message. The AI completed the test (added the != and .equals(None) TypeError assertions and renamed a variable that shadowed a builtin) and drafted this PR description.

PR code and description written by:

• Human
• AI

Reviewed before submission by:

• Human
• AI
• Not reviewed

• GitHub Issue: [Python][Parquet] Comparing metadata or schemas with None can crash Python #53500

[nleerollergh]

Testing evidence:

With the fix (branch as committed)
export PARQUET_TEST_DATA=$PWD/cpp/submodules/parquet-testing/data
cd python
python -m pytest pyarrow/tests/parquet/test_metadata.py -k equals_none -v
python -m pytest pyarrow/tests/parquet/test_metadata.py'

Without the fix, to confirm the regression test catches the bug
git checkout main -- python/pyarrow/_parquet.pyx
cd python && pip install -e . --no-build-isolation
python -m pytest pyarrow/tests/parquet/test_metadata.py -k equals_none
This ended in zsh: segmentation fault.

[nleerollergh]

Why the call crashes

In Cython, a parameter declared as Statistics other is checked to be a Statistics instance, but None is also allowed through unless you write not None. None is not a Statistics, so it has no .statistics attribute set up. Cython reads that attribute as a raw C++ field. The method body then runs deref(other.statistics.get()), and with other as None that reads garbage memory or a null pointer. Plain Python would raise AttributeError here. Cython does a direct struct access instead, so the process segfaults.

Why not None fixes it

With other not None, Cython adds a check at the very start of the call. If other is None, it raises TypeError immediately. That happens before the method body, so the dereference never runs.