test: fix flaky distributed vector build results test

[myandpr]

Summary

This fixes a flaky regression test added in #6220 (b80fbb3231cf58dd50e5670f9c56d309999bbd73).

The affected test is:

• test_distributed_vector_build_commits_multiple_segments_and_preserves_query_results

Recent failures showed up in both of these runs:

• https://github.com/lance-format/lance/actions/runs/23450834811
  main at 244c721504c6ef0b4c2f9700a342509976898d6e
• https://github.com/lance-format/lance/actions/runs/23460892697
  fix: preserve row last-updated metadata for rewrite-columns updates #6263

In those failures, different platforms / index variants failed:

• linux-arm / case_1_ivf_flat on main
• linux-build / case_2_ivf_pq on fix: preserve row last-updated metadata for rewrite-columns updates #6263

That points to an existing flaky test.

Root Cause

The test compared the exact Top-K _rowid results between:

• a single-segment index build, and
• a distributed multi-segment index build

However, the query path used by the test is ANN (ANNIVFPartition) under the default probing behavior. With partial probing, the candidate set can differ slightly between single-segment and multi-segment layouts, especially near the tail of Top-K. That makes exact _rowid equality too strict for this test and causes intermittent failures.

Fix

Make the test probe all IVF partitions before comparing Top-K row ids:

• add .minimum_nprobes(TWO_FRAG_NUM_PARTITIONS) to the test query

This keeps the existing strong assertion (ids_single == ids_split) but removes the probing-related source of nondeterminism.

Testing

Local verification:

• export PROTOC=/opt/homebrew/opt/protobuf@3/bin/protoc
• cargo test -p lance test_distributed_vector_build_commits_multiple_segments_and_preserves_query_results -- --nocapture

Observed result:

• case_1_ivf_flat ... ok
• case_2_ivf_pq ... ok
• case_3_ivf_sq ... ok

I also verified during debugging that with full probing enabled, repeated runs of the previously flaky ivf_flat / ivf_pq cases became stable.

[github-actions]

Review

LGTM. Clean one-line fix for test flakiness.

Setting minimum_nprobes(TWO_FRAG_NUM_PARTITIONS) (which equals the total number of IVF partitions) ensures all partitions are probed during the query, making Top-K results deterministic and eliminating the source of flakiness. The fix is applied inside collect_row_ids so it affects both the single and split dataset queries symmetrically.

No issues found.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[myandpr]

Hey @Xuanwo @wjones127 , could you take a look this PR when you have a chance?

This makes the distributed vector build test probe all IVF partitions before comparing Top-K row ids, which should stabilize the test and unblock CI.

[Xuanwo]

Thank you!

[myandpr]

@Xuanwo Thanks for the review — appreciate it!🙌