Scale-down: a worker drains in under a second
A worker pod deleted in the middle of a load test. On SIGTERM the poll loop stops, the in-flight flow finishes, and the process exits — no run is dropped.
Drain
pod deleted
under load
Finish in-flight run
0.42 s
Kubernetes removes the pod
worker exits — 0.49 s
gone in 1.7 s
In-flight flow completes
Teardown + "Worker stopped"
Pod object cleanup
Verified under load: 747/747 requests returned 200 and the flow-runs API showed zero failed runs. Queued runs stay in Redis and are picked up by the remaining fleet — aggressive scale-down policies are safe.