Fix flaky rollback-footprint assertion in TruncateOrphansAfterRecoveryIT

[Andrii Lomakin (andrii0lomakin)]

Summary

Fixes a flaky failure of TruncateOrphansAfterRecoveryIT.cleanReopenAfterRollbackSkipsPassAndLeavesNoOrphan that turned the integration pipeline red on develop (failed run, macOS arm / JDK 21). The fix is test-only; production code is unchanged.

The build runs with -Dmaven.test.failure.ignore=true, so the EnricoMi "Publish Test Results" step is the gate that fails the job when any test result is a failure.

Root Cause

The test asserted that the physical .pcl File.length() measured immediately after a rolled-back transaction equals the pre-TX baseline (expected: 1024L but was: 9216L). That is not a deterministic function of the rollback:

• A fresh cluster eagerly allocates its first data page when the first row is inserted.
• WOWCache may flush that buffered page to disk on its background thread at any point during the doomed transaction.
• The rollback restores the logical horizon but does not physically truncate an already-flushed page — a tolerated physical orphan, reclaimed lazily by the allocator or by the orphan-truncation pass on a dirty reopen (pinned by cleanReopenDoesNotRepairPreExistingPhysicalOrphan).

So File.length() after a rollback reflects background-flush timing, not the rollback outcome. On most platforms the eager page was still buffered when the assertion read the size (so it read the baseline); on macOS arm / JDK 21 the background flush had landed (9216 = 1024 + 8192). The same class of WOWCache lazy-flush timing bug was fixed for the commit direction in #1118.

Changes

core/src/test/java/.../TruncateOrphansAfterRecoveryIT.java (test-only):

• Replace the racy physical-size assertion with the deterministic invariant the rollback guarantees: after a clean reopen, SELECT count(*) FROM RolledBack reads 0. Read end-to-end (survives close + reopen).
• Guard against a vacuous pass: assert the RolledBack class still exists on reopen (it is created outside the doomed TX), so a zero count can never come from an absent class.
• Pin the count != 0 direction in the companion positive control committedAllocationGrowsPclSoRollbackAssertionIsNonVacuous, which now also asserts the same count(*) query reads GROWTH_ROW_COUNT against a committed population.
• Sync the method/helper Javadoc and the surviving size-assertion message (which only proves the reopen is a physical no-op, not that the rollback left no flushed page).

The storage-level contract (WAL not replayed, orphan pass skipped, .pcl size stable across the reopen) stays covered by the three unchanged assertions. The primary proof that a rollback discards its footprint remains the unit test EndAtomicOperationRollbackSkipsCommitTest.

Known unfixed: a second, unrelated flake in the same run

The same failed run also failed LocalPaginatedStorageRestoreFromWALIT.testSimpleRestore on Linux arm / JDK 21 with RecordNotFoundException: #19:121 not found during the post-restore DatabaseCompare. This is a genuine, rare WAL-restore/SI divergence (the WAL-restored copy is missing a delete the live DB applied), not a test-harness artifact — the same bug class has been patched repeatedly (YTDB-510, YTDB-545, and a prior @Ignore). It is deliberately not addressed here: masking it with a retry or @Ignore would hide a real durability divergence, and a sound root-cause fix is a separate, deep storage investigation. Both failures are flaky and unrelated to the triggering commit (YTDB-651, which touches only SQL prefilter / MATCH executor code); the same tests passed on JDK 25 across all architectures and on Linux x86 / Windows under JDK 21.

Motivation

CI failure on develop: https://github.com/JetBrains/youtrackdb/actions/runs/26942924852

Test plan

• cleanReopenAfterRollbackSkipsPassAndLeavesNoOrphan passes locally (-Dyoutrackdb.test.env=ci)
• committedAllocationGrowsPclSoRollbackAssertionIsNonVacuous passes (confirms the committed run persists exactly GROWTH_ROW_COUNT rows)
• Full TruncateOrphansAfterRecoveryIT class (34 tests) passes
• spotless:check clean; all core test sources compile
• Dimensional review completed (test-behavior, test-completeness, test-structure, code-quality); all should-fix findings addressed

[gemini-code-assist]

Note

Gemini is unable to generate a review for this pull request due to the file types involved not being currently supported.

[github-actions]

Test Count Gate Results

Tolerance: 5% drop allowed per module

Overall: ✅ 29108 tests (baseline: 29108, +0)

Module	Baseline	Current	Change	Status
core	18560	18560	+0	✅
docker-tests	1891	1891	+0	✅
embedded	1931	1931	+0	✅
examples	6	6	+0	✅
gremlin-annotations	30	30	+0	✅
jmh-ldbc	39	39	+0	✅
server	5504	5504	+0	✅
tests	1147	1147	+0	✅

[github-actions]

Coverage Gate Results

Thresholds: 85% line, 70% branch

Line Coverage: ✅ No coverable lines in diff

Branch Coverage: ✅ No branches in changed lines