Re-enable LMDB caching tests #414

Activity

1. kriszyp

   added this to the v5.0 milestone on May 21, 2026
2. kriszyp

   added a commit that references this issue on Jun 18, 2026

   test(blob): fix startPreCommitBlobsForRecord.complete() assertion for…

   ...

   3b3d138
3. kriszyp

   added a parent issue on Jul 7, 2026

   • [Epic] Test-suite debt: skipped/disabled/flaky tests #1669
4. kriszyp

   modified the milestones: v5.0, v5.3 on Jul 23, 2026
5. kriszyp

   set

   Priority

   to

   P3

   on Aug 7, 2026
6. kriszyp

   added the

   Task

   issue type on Aug 7, 2026
7. kriszyp commented on Sep 16, 2026

   

   kriszyp

   on Sep 16, 2026

   MemberAuthor

   More actions

   Three concrete reproductions from the week of 2026-09-08, in case they help turn this from "cannot reproduce outside CI" into something with a thread to pull. All three were re-fetched on 2026-09-16 and the logs were still retrievable.

   job	branch @ commit	Node	signature
   104087005044	fix/integration-flake-triage @ 21ad982d5	22	Segmentation fault (core dumped), exit 139, no mocha failing block
   104198556436	deps/s3-optional-peer @ 1f524205f	24	identical signature
   104232294559	deps/s3-optional-peer @ dbea34171	22	The action 'Unit tests: lmdb' has timed out after 4 minutes — a hang, not a crash

   Both segfaults land in a 9–20 second gap right after the last test completes, which points at env-close / final-commit rather than any individual test body. No core dump or backtrace was captured, so that is a plausible location, not a proven one.

   If it is the teardown path, the freelist/commit family in #685 and #686 is the obvious suspect — mdb_page_alloc reaching MDB_NOTFOUND via the fork-local me_freelist_start cache, and rejectCommit's unhandled rejection. I want to be clear that I have not established that link; it is a hypothesis worth one targeted run with a core dump enabled, not a conclusion.

   Two constraints on any archaeology here, worth writing down because they are not obvious:

   • Unit tests: lmdb only became a distinct step on 2026-09-14 (Split unit-test.yml's Run tests step into one step per suite #2588, a17a3eeaa7, which split unit-test.yml one step per suite). Any sweep that filters on that step name is structurally blind before that date.
   • This is not the same failure as Flaky: Unit Test (Node.js v24) SIGSEGV (exit 139) in schema-change / index-rebuild test group — v24-only #1380, which is a v24-only SIGSEGV in the schema-change / index-rebuild step and self-describes as distinct.

   Severity framing: LMDB is deprecated and RocksDB is the supported engine, so this is CI-maintenance cost rather than product risk — which is probably the right reason to keep it at whatever tier it has rather than escalating it.

   Surfaced by finding-triage from a main-red investigation.