Laravel 13.37 Fixes Unique Job Locks Stuck After a Failed Push

Unique Lock Fixes in Laravel 13.36 and 13.37 #

A Failed Push Left the Lock Held Forever #

PR #62010 describes the most common case. PendingDispatch acquires the unique lock in shouldDispatch() and then pushes the job, but nothing released the lock when the push threw. Once the queue was back, the job could not be dispatched again until the lock expired, and with the default uniqueFor of 0 it never does:

class SyncInventory implements ShouldQueue, ShouldBeUnique
{
    //
}

SyncInventory::dispatch(); // RedisException: Connection refused
// later, with the queue back up
SyncInventory::dispatch(); // skipped, the lock is still held

From 13.37.0 the lock is released when dispatching throws, the exception is rethrown, and the unique job information added to the context is removed. The release goes through UniqueLock::release(), so it only releases the lock that dispatch owns. 13.37.0 applies the same fix to unique queued listeners (#62045).

Rolled-Back Transactions and Unique Queued Listeners #

In 13.36.0, PR #61974 fixes the same symptom for listeners that implement ShouldQueueAfterCommit and ShouldBeUnique. When the event was fired inside a transaction that rolled back, the listener was never queued but its lock stayed, so the next event for the same order was skipped. An earlier fix, #55420, already released the lock on rollback for unique jobs, but the check only looked for ShouldBeUnique, while queued listeners are pushed as a CallQueuedListener that reports uniqueness through shouldBeUnique(). The PR adds the check the queue handler already used.

The onOneServer and withoutOverlapping Race #

PR #62007, which closes issue #50330, explains why a task using both guards was occasionally skipped on every server: CacheEventMutex::exists() checked whether a task was running by acquiring the event lock and releasing it, so for a moment other servers saw the task as overlapping. The fix uses Lock::isLocked(), which reads the lock's state without acquiring it. The author notes that the failure is silent and was seen about 1 to 3% of the time on one project, so it may be under-reported. An earlier attempt, #54430, did not handle locks kept in a different cache store, such as the database cache_locks table or a Redis lock_connection.

Upgrading and Protecting Older Versions #

1. Upgrade to Laravel 13.37.0 or later; it carries the 13.36.0 fix as well.
2. Until then, give unique jobs and listeners a finite uniqueFor, so a lock left behind by a failed push or a rollback expires instead of blocking the job forever.
3. Look for unique jobs that stopped running after a queue outage: their locks may still be held in the cache store and can be cleared there.
4. If scheduled tasks use both onOneServer() and withoutOverlapping(), check run monitoring for skipped runs; after 13.37.0 the race no longer applies.
5. Test suites that call Queue::fake() more than once in one process now see unique queued listeners pushed each time.