mirror of
https://github.com/windmill-labs/windmill.git
synced 2026-09-10 16:05:58 +00:00
Ruben FiszelandClaude Opus 4.8
fa3644281f
handle_zombie_flows() joined worker_ping on the flow's queue-row worker
(q.worker) only. In nested/subflow/forloop cases that worker is frequently
NOT the one that performed the flow's final state transition, so all OOM
detection ran against a healthy bystander and the diagnostic wrongly told
on-call to chase a deadlock / SIGQUIT.
Change 1: when q.worker itself looks healthy, additionally search worker_ping
for a plausible culprit — a different worker on the same pod (worker_instance)
or worker group whose last ping clusters around the flow's last_ping and has
since gone silent (the signature of a worker OOM-killed mid state-transition),
picking the one closest in time to the transition. If found, the message names
that worker and routes to the OOM explanation. The extra query runs only on the
rare zombie path and fails soft (warn + fall back) so it never blocks the cancel.
Change 2: reorder the healthy-q.worker hint to lead with "check whether a
different worker on the same pod/group was OOM-killed around {last_ping}",
keeping deadlock/SIGQUIT as the secondary possibility.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>