16. Nodes claim only job types they handle
Date: 2026-10-05
Status
Accepted
Context
Nodes claimed every due job and failed an attempt for a type they had no handler for. During a
rolling deploy, nodes of the old version took jobs of types only the new version handles and used
up their attempts until they ended FAILED. db-scheduler leaves such “unresolved” tasks to other
instances.
Decision
The claim only selects jobs whose job_type is one of the node’s registered handler types
(job_type = any(?)). job_type becomes the last key column of the polling index
(priority desc, scheduled_at, job_type), so the condition is checked on index entries in claim
order, without reading the rows of skipped jobs (measured: 36 instead of about 5,400 buffers with
100,000 due jobs of an unhandled type ahead in claim order). No separate index per type: claiming
across types in priority order needs the order of the existing index.
Consequences
- Jobs of a type no node handles stay
SCHEDULEDinstead of failing; there is no automatic clean-up or warning for them yet. - Many due jobs of unhandled types ahead in claim order still cost index reads on every poll.