Jobs (3.x): two runs claim and run the same due jobs; the slower one fails with "Cannot read properties of null (reading 'log')" — please backport #17441 #18565

Describe the Bug

In 3.90.2 (Postgres adapter), two job-run calls a moment apart can claim and run the same due jobs. Claiming is a find followed by one update per row (@payloadcms/drizzle/dist/updateJobs.js), outside any transaction, so both runs see the same jobs with processing: false and both mark and run them.

With the default deleteJobOnComplete: true, the faster run completes and deletes each job. The slower run then fails to insert its log row (insert into "payload_jobs_log" … Failed query), reads the job back as null, and throws TypeError: Cannot read properties of null (reading 'log'). The run endpoint answers 500. Every task also runs twice: a scheduled publish publishes the document twice.

This is the race in #16043. It was fixed by #17441 ("feat!: support multiple workers on the same queue"), which was merged into main (4.0) only. The latest 3.x (3.90.2) still has the find-then-update claim. Could the fix, or a non-breaking part of it, be backported to 3.x?

Link to the code that reproduces this issue

The reproduction below uses only Payload's own scheduled publish and run endpoint, with no project code.

Reproduction Steps

1. Payload 3.90.2 with @payloadcms/db-postgres 3.90.2, a collection with versions: { drafts: { schedulePublish: true } }, and jobs.access.run allowing a bearer token.
2. Schedule a publish of three documents for the same minute (the admin's "Schedule Publish", or payload.jobs.queue with the schedulePublish task).
3. Once they are due, call GET /api/payload-jobs/run twice, about a second apart. Our tests do this from three parallel workers.
4. Observed, in about 1 run in 3:
   • One call answers 200, and the other answers 500 with {"message":"An unknown error has occurred.","noJobsRemaining":true}.
   • The server logs Error running job Task: schedulePublish … Failed query: insert into "payload_jobs_log", then There was an error running jobs: TypeError: Cannot read properties of null (reading 'log').
   • The audit trail (an afterOperation hook) records two publishes for each document.

Expected: each due job is claimed by exactly one run, as it is in 4.0 after #17441.

Our workaround, for anyone on 3.x: we closed the REST run endpoint (jobs.access.run: () => false) and run jobs only from our own route. That route takes a Postgres transaction-level advisory lock (pg_try_advisory_xact_lock) on a dedicated connection; a transaction-level lock because Neon's pooler is in transaction mode. We also wrap the schedulePublish handler so that a publish which is already done is skipped.

Which area(s) are affected?

area: core, db: postgres

Environment Info

payload: 3.90.2
@payloadcms/db-postgres: 3.90.2
@payloadcms/drizzle: 3.90.2
next: 16.3.8
node: 22 (CI), 24 (local)
Postgres: Neon (pooled connection, transaction mode)
OS: Linux (CI), Windows 11 (local)