Repository navigation
Disable advisory lock transactions by default - #5526
Merged
Merged
Conversation
This was referenced Sep 29, 2026
Closed
jeremydmiller
added a commit
that referenced
this pull request
Oct 4, 2026
Closes #5567. Reported with a reproduction, a dump, and three candidate fixes, all useful. Stopping a HotCold daemon holding many projection locks could stall 60 or 120 seconds before the locks were released, silently, on the 9.41+ defaults. Reproduced on master: 60091 ms on iteration 10 of one run, 60317 ms on iteration 1 of the next, nothing in two others. The cause is in Patched.DistributedLock.Core 1.0.11, both halves visible in its shipped IL. MultiplexedConnectionLock.Handle.DisposeAsync disposes the monitoring handle BEFORE the ReleaseAsync that needs the shared connection for pg_advisory_unlock, and AcquireConnectionLockAsync fires its wake-up only while registrations remain -- so the monitor worker commits to monitoring from a count the releaser has since decremented, the last release on a connection wakes nothing, and its own 2s retry loop re-evaluates that same false condition until the one-minute window expires. Two such connections make it 120 s. It is NOT a failure to interrupt the wait: NpgsqlConnection.WaitAsync honours its token promptly, including the pre-cancelled-and-disposed-CTS shape FireStateChangedNoLock produces (0 ms) and cancellation mid-wait (1003 ms). The interruption is never requested. And no bump fixes it -- Postgres publishes only 1.3.2/1.3.3 and Core only 1.0.10/1.0.11, both unlisted, Weasel already references the newest, and the fork is ahead of upstream's 1.3.1. So AdvisoryLock moves out of Weasel.Postgresql into Marten.Events.Daemon.Coordination, where the release path can be fixed without a two-repository release chain. Kept free of anything Marten-specific; the intent is to move it back to Weasel once it has settled. DisposeAsync now releases concurrently and bounds the wait with Projections.StopAndDrainTimeout, leaving anything outstanding to finish in the background, and logs the overrun -- this path previously logged nothing unless it threw, which is why a minute-long stop came with no explanation. Measured over 30 iterations per configuration, the formerly stalling config is min 56 / median 792 / p90 980 / max 988 ms. Releasing concurrently does not cost the common case, which was the risk worth checking: all the monitoring registrations drop at once, so the intermediate wake-ups a sequential release triggered never happen, and the fire when the count reaches zero turns out to be what unblocks it either way. Not multiplexing also fixes it -- one registration per monitor always reaches zero -- and was rejected: 24 locks went from 1 backend connection to 24. A non-positive or infinite ReleaseTimeout opts OUT of the bound rather than meaning "give up immediately", matching how JasperFx reads DaemonSettings.StopAndDrainTimeout. Run against the unguarded code only the negative case fails; zero and infinite pass either way because releases finish in milliseconds, which the test file records so nobody reads them as coverage they aren't. Docs: #5526 removed the only mention of UseMonitoredAdvisoryLock while flipping the transaction default, leaving it undocumented anywhere, and the text it removed had gone stale -- on a data source Marten owns the monitor no longer runs SELECT pg_sleep(60), it waits client-side and the backend sits idle, which is why the stall was invisible in pg_stat_activity as well as in the logs. Marten.Events.Daemon.Coordination.AdvisoryLock is a source-level CS0104 ambiguity for code importing both that namespace and Weasel.Postgresql. Not a binary break; the one in-repo case drops the Weasel import. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01481eiEZg1Bv6pu4DrgPRg3
This was referenced Oct 6, 2026
Closed
This was referenced Oct 16, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.