Pre-PR Validation Gate (session: cron:clickhouse-ci-task-worker:20260402-081500)

a) Deterministic repro? Yes. SET query_plan_merge_expressions = 0; EXPLAIN header = 1 SELECT n.a FROM t_nested ARRAY JOIN n ORDER BY n.a; deterministically produces different output from the reference.

b) Root cause explained? The test's EXPLAIN header=1 output shows merged Expression step names like (Before ORDER BY + Projection). When query_plan_merge_expressions=0 is randomized by CI, these steps appear separately as Expression (Before ORDER BY) + Expression (Projection), and the [lifted up part] annotations from expression lifting also disappear. query_plan_enable_optimizations=0 has the same effect as the master switch.

c) Fix matches root cause? Yes — pins exactly the two settings that control expression merging: query_plan_enable_optimizations=1 (master switch) and query_plan_merge_expressions=1 (specific optimization).

d) Test intent preserved? Yes — the test verifies ARRAY JOIN column pruning. The pinned settings only affect EXPLAIN formatting (step merging), not the actual column pruning behavior being tested. All assertions about which columns are read remain unchanged.

e) Both directions demonstrated? Yes — without the fix + query_plan_merge_expressions=0: EXPLAIN output differs (steps split). With the fix: 50/50 passes with full CI randomization.

cc @novikd — could you review this? It pins query_plan_enable_optimizations=1 and query_plan_merge_expressions=1 in the ARRAY JOIN column pruning test to prevent EXPLAIN output differences when CI randomizes these settings.

Workflow [PR], commit [e3f9b38]

Summary: ❌

AI Review

Summary

This PR stabilizes tests/queries/0_stateless/04039_prune_array_join_columns.sql by pinning query_plan_enable_optimizations=1 and query_plan_merge_expressions=1, so EXPLAIN header=1 output is no longer affected by CI randomization. The change is minimal, targeted, and matches the flaky behavior root cause; I found no correctness, safety, or performance issues.

ClickHouse Rules

Final Verdict

• Status: ✅ Approve

Closing: this flakiness is caused by unmerged PR #100874 which randomizes query_plan_merge_expressions — a setting that per @KochetovNicolai should always be enabled and is only useful for debugging. It is not randomized on master, so these tests don't fail on master. The fix belongs in PR #100874 (don't randomize this setting), not in pinning it across dozens of test files. Apologies for the noise.