[MOD-19372] Fix flaky SVSTieredIndexTest.updateVectorsDuringUpdateJob

[nonirosenfeldredis]

Jira: MOD-19372

Problem

SVSTieredIndexTest/1.updateVectorsDuringUpdateJob (uncompressed, multi-value) failed intermittently in CI (merge-queue run 37461178846) at the final per-label top-1 check:

tests/unit/test_svs_tiered.cpp:498: Failure
  id  Which is: 98
  i   Which is: 100

Root cause

Not data loss. Reproduced locally under heavy CPU load (about 6 failures in 1,100 runs with 24 parallel processes; 40/40 pass serially) and diagnosed the failing assertion: the label's own vector is stored in SVS at distance 0 from the query, the flat buffer is empty, and the same query with a window of 3000 returns the right label. The node is under-connected in the Vamana graph (same family as MOD-18994). How the mock workers batch the update-job drain depends on timing, so how under-connected a node ends up differs from run to run, and the test's search window of 250 occasionally did not reach it.

Fix

Test-only: remove the top-1 search at the end of the test. It checks graph reachability, which varies with timing and is not what this test covers. What remains still covers the update-during-update-job window: every label is held in exactly one tier, and the backend holds exactly the replacement vectors (indexSize() - numMarkedDeleted == n * new_per_label), which fails if a job carries a replaced vector into the backend after its delete.

The underlying graph-construction behavior (MOD-18994) is not changed here.

Test plan

• SVSTieredIndexTest/*.updateVectorsDuringUpdateJob builds and passes (3 variants, ASan debug build).
• clang-format 18.1.8 clean.
• Not run locally: the rest of test_svs and the other tiered suites; relying on CI.

🤖 Generated with Claude Code

---

Note

Low Risk
Test-only change; no production code or indexing behavior is modified.

Overview
Test-only fix for intermittent CI failures in SVSTieredIndexTest.updateVectorsDuringUpdateJob (MOD-19372).

The PR drops the final block that ran top-1 KNN searches (with a widened search window of 250) on a few sample labels for uncompressed indexes. Those checks exercised approximate graph reachability after batch drains, which can vary with worker timing and under-connected Vamana nodes (same family as MOD-18994)—not the update-during-job correctness this test is meant to prove.

What still runs: updates while update jobs are in flight, then assertions that every label exists in exactly one tier and that indexSize() - getNumMarkedDeleted() equals n * new_per_label (the check that catches replaced vectors wrongly carried into the backend). No production or SVS graph behavior changes.

^{Reviewed by Cursor Bugbot for commit f2bed41. Bugbot is set up for automated code reviews on this repo. Configure here.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.53%. Comparing base (ab6bb4f) to head (f2bed41).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1076      +/-   ##
==========================================
- Coverage   97.54%   97.53%   -0.02%     
==========================================
  Files         142      142              
  Lines        9405     9405              
==========================================
- Hits         9174     9173       -1     
- Misses        231      232       +1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.