[hcrosse]

Closes #4277

Rationale for this change

delete(), overwrite(), upsert() and dynamic_partition_overwrite() raise ValueError: Cannot create expression literal from NaN when they touch a NaN identity partition. _build_partition_predicate builds EqualTo for each non-NULL partition value, and expression literals reject NaN. I changed _build_partition_predicate to emit IsNaN for NaN partition values next to the existing IsNull branch.

This PR depends on #4280. On main, Record compares two NaN partition values as unequal, so commit validation misses a concurrent position delete in a NaN partition. Today the replacement raises before validation runs. With only this change, a stale file replacement would commit over that delete and bring the deleted row back. This PR includes both commits, and only the second one belongs to it.

Dynamic overwrites with NaN keys used to raise. They can now reach two existing bugs, both of which need a file written under an older partition spec.

• With both NaN and NULL keys, a dynamic overwrite can drop a NULL row that matches none of the incoming keys. This is Table.delete() with an OR filter deletes the wrong rows when a column is NULL #4184. A hand-written delete() filter already loses the row on main. The open Keep rows where a delete filter evaluates to null #4219 keeps the row for that filter. I did not test the dynamic-overwrite case against it.
• With two partition columns, a dynamic overwrite can drop a whole old file, including rows that match none of the incoming keys, because PyIceberg writes no NaN counts. This is Data files written by PyIceberg have no nan_value_counts, so delete() drops NaN rows that do not match the filter #4130. Finite keys already reach it on main.

upsert() with NaN in a join column still raises, because create_match_filter builds literals from the join keys. That path is separate from this change.

Are these changes tested?

Yes. I added three tests.

• test_replace_files_in_nan_identity_partition covers whole-file deletion, partial rewriting, overwrite and dynamic overwrite in a NaN identity partition, on the memory and SQL catalogs.
• test_build_partition_predicate_with_nan_and_null checks the predicate for NaN, NULL and finite values.
• test_file_overwrite_validates_concurrent_position_delete_in_nan_partition stages a file replacement, then commits a partition-scoped position delete from another table handle. It checks that the replacement fails validation and the deleted row stays gone, on the memory and SQL catalogs. PyIceberg cannot write delete files, so the helper commits the delete through overwrite() with _OverwriteFiles.new_manifest_writer patched to write a delete manifest.

All three fail on main with the NaN literal error and pass with both commits. Without the dependency, the concurrent-delete test fails because the stale replacement commits.

I also ran upsert() on a row in the NaN partition and delete(AlwaysTrue()). Both raise on main and succeed with the fix. These cases are not in the committed tests.

The full unit suite with both commits gives 4482 passed, 4 failed and 5 skipped, versus 4472 passed, 4 failed and 5 skipped on main. Three of the ten new passes come from the dependency and seven from this PR. The same four Hive Kerberos tests fail on both because the hive-kerberos extra cannot be installed in the test environment. prek run -a passed.

Are there any user-facing changes?

Yes. delete(), overwrite(), upsert() and dynamic_partition_overwrite() now work on NaN identity partitions, apart from the cases above. There is no API change. Please add the changelog label.

[Copilot]

🟡 Changes recommended

The equality fast path repeats comparisons for unequal nested Records, causing exponential work with nesting depth instead of a single linear pass.

1 open issue

• Avoid repeating comparisons of unequal nested Records · New

What changed in this PR

This PR fixes partition predicate construction for NaN identity partitions, allowing affected delete and replacement operations to proceed. It includes the dependency that makes NaN partition records compare equal for delete lookups and conflict validation.

Changes:

• Emit IsNaN for NaN partition values while preserving NULL and finite-value handling.
• Treat corresponding NaN floats as equal in Record.
• Add regression coverage for file replacement, partition lookups, and concurrent-delete validation.

File	Description
pyiceberg/​table/​__init__.py	Build NaN-aware partition predicates.
pyiceberg/​typedef.py	Add NaN-aware record equality.
tests/​test_typedef.py	Cover NaN record equality and set deduplication.
tests/​table/​test_validate.py	Check concurrent deletion conflicts in NaN partitions.
tests/​table/​test_init.py	Verify predicates for NaN, NULL, and finite values.
tests/​table/​test_delete_file_index.py	Verify delete lookup using NaN partition records.
tests/​table/​test_commit_retry.py	Reject replacements conflicting with concurrent position deletes.
tests/​catalog/​test_catalog_behaviors.py	Exercise deletes and overwrites in NaN identity partitions.

_{Review effort: Balanced}

[hcrosse]

This comes from the commit this PR includes from #4280 and is fixed there. This PR picks up the fix when it merges main after that PR lands.