Abandon (not crash) evicted flow runs when Kubernetes owns retries#22622
Open
chuqCTC wants to merge 1 commit into
Open
Abandon (not crash) evicted flow runs when Kubernetes owns retries#22622chuqCTC wants to merge 1 commit into
chuqCTC wants to merge 1 commit into
Conversation
chuqCTC
force-pushed
the
fix/22620-abandon-sigterm-backofflimit
branch
3 times, most recently
from
July 24, 2026 21:05
1552afd to
4bdc7c1
Compare
chuqCTC
marked this pull request as ready for review
July 24, 2026 21:07
chuqCTC
requested review from
chrisguidry,
desertaxle and
zzstoatzz
as code owners
July 24, 2026 21:07
chuqCTC
force-pushed
the
fix/22620-abandon-sigterm-backofflimit
branch
from
July 24, 2026 21:08
4bdc7c1 to
c368e70
Compare
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 4bdc7c19bf
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
Merging this PR will not alter performance
Comparing 
Footnotes
|
Closes PrefectHQ#22620. With a Kubernetes work pool at backoffLimit > 0, a SIGTERM (eviction) made the in-pod engine finalize the run Crashed via handle_crash, but Kubernetes then retried the Job and re-ran the flow, so the run left a terminal state and executed twice. Add an "abandon" SIGTERM behavior: when backoffLimit > 0 the worker sets PREFECT_FLOW_RUN_EXECUTE_SIGTERM_BEHAVIOR=abandon, and the engine bubbles the raw SIGTERM without proposing a state. The run stays non-terminal while Kubernetes retries the pod; if retries are exhausted the worker's non-zero-exit crash proposal finalizes it. Prefect reschedule and k8s Job backoff remain mutually exclusive. A raw SIGTERM can reach an async flow as a cancellation rather than a TerminationSignal, so the bubble is gated on a latch set by the SIGTERM handler to avoid abandoning ordinary cancellations (e.g. flow timeouts).
chuqCTC
force-pushed
the
fix/22620-abandon-sigterm-backofflimit
branch
from
July 24, 2026 21:26
c368e70 to
b6df6d9
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
closes #22620
The bug
On a Kubernetes pool with
backoffLimit > 0, an evicted pod gets a SIGTERM and the engine marks the runCrashed. But Kubernetes then starts a fresh pod for the same Job and runs the flow again, so a run that was already terminal comes back and runs twice.The fix
Add an
abandonSIGTERM mode. WhenbackoffLimit > 0, the worker setsPREFECT_FLOW_RUN_EXECUTE_SIGTERM_BEHAVIOR=abandon, and in that mode the engine exits on SIGTERM without setting any state. The run stays running while Kubernetes retries the pod. If every attempt is evicted, the worker still marks the runCrashedonce the retries are used up.Prefect's own reschedule (
backoffLimit == 0) and Kubernetes Job retries stay mutually exclusive, as before.Worth noting
backoffLimitnow behaves likeabandonrather than crashing. Kubernetes defaults it to 6, so retries really do happen.reschedulemode for async flows, where a delayed SIGTERM could arrive as a cancellation and crash the run instead of leaving it inAwaitingRetry.