Pre-PR Validation Gate

Session: cron:clickhouse-ci-task-worker:20260402-011500

a) Deterministic repro? ✅ Yes — SET query_plan_merge_expressions = 0 triggers the failure 100% of the time. Every query in the test returns brute-force L2Distance results instead of HNSW index results.

b) Root cause explained? ✅ tryUseVectorSearch() in useVectorSearch.cpp walks the query plan expecting exactly: LimitStep → SortingStep → ExpressionStep → [FilterStep] → ReadFromMergeTree. When query_plan_merge_expressions = 0, the plan has unmerged ExpressionSteps (e.g., one for the ORDER BY distance function, one for SELECT list). The pattern matcher at line 76-78 finds the first ExpressionStep, then at line 82-83 expects ReadFromMergeTree or FilterStep but finds another ExpressionStep instead → returns no_layers_updated → vector search optimization not applied → brute-force scan.

c) Fix matches root cause? ✅ Pinning query_plan_merge_expressions = 1 ensures expressions are merged before the vector search optimizer runs, preserving the expected plan structure.

d) Test intent preserved? ✅ The test verifies vector search rescoring and prewhere interaction. query_plan_merge_expressions controls a general expression merging optimization unrelated to vector search. Pinning it preserves the test's ability to catch vector search bugs.

e) Demonstrated in both directions? ✅

• Without fix + merge_expressions=0: returns 16, 19, 12, 11 (brute force) — FAIL
• With fix: 50/50 passes with full CI randomization — PASS

cc @rschu1ze — could you review this? It pins query_plan_merge_expressions=1 in the vector search rescoring test because tryUseVectorSearch() needs merged expression steps to recognize the LimitStep→SortingStep→ExpressionStep→ReadFromMergeTree plan pattern. When expressions aren't merged, the optimizer silently falls back to brute-force scan, producing wrong results.

Workflow [PR], commit [3107817]

Summary: ❌

AI Review

Summary

This PR stabilizes tests/queries/0_stateless/02354_vector_search_rescoring_and_prewhere.sql by pinning query_plan_merge_expressions = 1, preventing random-plan variation from switching vector-search optimization off and changing expected results. The change is minimal, targeted, and appropriate for a flaky-test fix. I did not find correctness, safety, performance, or compliance issues in the diff.

ClickHouse Rules

Final Verdict

• Status: ✅ Approve