[bryancluo]

Summary

Adds a skip_null option to fragment update_columns. With it set, a null in the update stream keeps the existing value of the matched row instead of overwriting it, so callers can send partial updates without reading the old values first.

Semantics:

• Only top-level nulls are skipped. A non-null struct still replaces the old struct as a whole, including its null children.
• An update row whose columns other than the join key are all null changes nothing, so it is treated as unmatched. It is left out of matched_offsets, and its _row_last_updated_at_version does not change after a rewrite_columns commit. If every update row is like this, the call is a no-op, the same as an empty update stream.
• The default (false) keeps the current behavior.

Implementation

• UpdateJoinOptions::with_skip_null / skip_null() carry the flag into the internal UpdateColumnsOptions.
• All-null update rows are filtered out of the RHS before strategy selection, so they never join on either path.
• Hash path: HashJoiner::with_skip_null computes the RHS null masks once. collect_with_fallback then points null cells at the left row in the interleave indices, so each column is still materialized only once.
• Sort-merge path: SortedPatchCursor::merge_batch does the same per-column index rewrite on its patch batches.

Bindings

• Python: keyword-only skip_null=False on LanceFragment.update_columns.
• Java: a Fragment.updateColumns(stream, leftOn, rightOn, skipNull) overload. The 3-argument method delegates with skipNull = false.

Tests

• Rust update_join::tests::test_update_skip_null: hash × sort-merge × skip_null, with two fragments and an RHS split across batches. It covers a top-level null string, a top-level null struct, a non-null struct with a null child, an all-null row, an unmatched row, matched_offsets, and _row_last_updated_at_version after commit.
• Rust test_empty_rhs_is_noop: new case for an RHS made only of all-null rows.
• Python test_fragment_update_columns_skip_null and Java UpdateTest.testUpdateColumns are parameterized on skip_null, including offsets and version metadata.

🤖 Generated with Claude Code

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[lance-gatekeeper]

✅ Gate recommendation: approve.

The opt-in null handling fits the existing fragment rewrite contract. The focused Rust tests passed for both join strategies, including struct replacement, matched offsets, committed row-version metadata, and all-null no-op inputs.