Harden the legacy-row partial-index plan test: fixed anchor, size guard, plan in the failure message

[erikdarlingdata]

Summary

TheShippedCheck_PlansOnThePartialIndex_AndAnswersExactly failed once on a dev push's Build run on 2026-10-04 at about 00:08Z, at Assert.NotEmpty(indexReads): a scan of query_store_stats was found, but none read the partial index. It passed on re-run, and the assert printed no plan. It failed the same way on another PR's Build run about 15 minutes later, a PR that changes only web JavaScript. Both failures fell in the first half hour after 00:00 UTC, after a day of green runs.

The test seeded to the wall clock, so the one uncompressed chunk was "today since 00:00 UTC". Just after midnight that is one pass, about 60 rows on one heap page.

I could not reproduce the failure. With the old shape (only the time of day varied), anchors 00:01, 00:03, 00:08 (14 runs), 00:31, 00:38, 01:08, 01:38, 02:08, 03:00, 06:00, 12:00 and 23:55 all passed. At 00:08 the planner picked an Index Only Scan on the partial index for the one-pass chunk:

->  Index Only Scan using _hyper_*_chunk_idx_query_store_stats_server_time_null_start on _hyper_*_chunk s_4  (cost=0.12..1.73 rows=1)

With index and bitmap scans disabled, the competing Seq Scan on that chunk costs 0.00..2.05. So the near-empty chunk is a close call (1.73 vs 2.05), which is a plausible way for a small difference in statistics, visibility-map state or the platform to flip the plan on CI. That is an inference, not something I reproduced. The partial index is empty in this test (no row has a NULL start), so its cost stays about constant as the chunk grows, while the Seq Scan's cost grows with the chunk's pages. A full-day chunk therefore widens the gap instead of sitting on it. The earliest time of day at which the old shape passes is therefore not a meaningful number: it passed at every time I tried.

What changes

Tests only, in QueryStoreLegacyRowIndexLiveTests.cs:

• The seed ends at a fixed anchor (23:50 UTC on a fixed past day) instead of DateTime.UtcNow. The newest, uncompressed chunk always holds a full day of passes (about 2,880 rows), whatever time the test runs. Nothing else in the test compared against the real "now".
• A guard asserts that the newest chunk holds at least 1,000 rows, with a message, so a later seed change cannot bring the tiny-chunk shape back.
• Assert.NotEmpty(scans) and Assert.NotEmpty(indexReads) became Assert.True(..., message + "\n" + plan), so the next failure shows the plan.
• No enable_seqscan = off, no forced plans, no VACUUM. The test still asserts that the shipped check plans on the partial index and never heap-filters the NULL test. I did not try VACUUM, since there was no failure to stabilise.

Test plan

• Old shape, anchors listed above: pass, no RED. (This is the limit: I have no RED.)
• With the fix: the class passes 3 runs in a row, and with temporary anchor inputs of 00:05, 06:00, 12:00 and 23:55 (reverted before commit).
• TuningHourlyCreateLiveTests: 5 of 5 pass. LivePostgresCollectionHygieneTests cannot load WPF assemblies on this machine.
• Full suite without DARLING_TEST_PG: 22046 total, 476 failed, and every failing test name is already on the known local-failure list for dev (the comm -23 difference is empty).
• --no-incremental build: 0 warnings.

Limits

• The root cause is not proven; see the Summary. If the test fails again, the message now includes the plan, which will settle it.
• The size guard runs only when TimescaleDB is present, as the chunk split does.

CHANGELOG

SECTION: None
ENTRY: None: tests only