Repository navigation
Replies: 1 comment 1 reply
|
The scheduling example is the one I’d want in a Slack alert. If the GPU nodes are short on memory and the other nodes don’t match the workload, seeing both reasons saves a trip to the cluster. A plain timeout leaves the app team guessing whether to change the deployment or ask for capacity. One case I’d put in the RFC tests: the old ReplicaSet still has a pod in ImagePullBackOff, but the new ReplicaSet is stuck on FailedScheduling. Which reason makes it into |
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
When a Kustomization health check fails, Flux reports which objects aren't ready:
The actual cause is almost always one level further down, on the pods of the workload, and
belongs to a small set:
Anyone with cluster access finds this with
kubectl describein a minute. Application teamson shared platforms often don't have that access, and every Flux notification (Slack, commit
status, ...) carries only the summary above. Platform teams end up answering "my deployment
failed, why?" by hand.
We built an in-cluster component that reads the failed objects' workloads, pods and warning
events and condenses them into one line, for example:
It works, but it duplicates logic that would serve every Flux user better inside
kustomize-controller and helm-controller. Then the cause would be in the
Readymessage andin events, so it flows to all notification providers.
Would maintainers be open to this, opt-in behind a feature gate? We'd contribute the
implementation, and we're looking for a sponsor for the RFC, drafted in #6187. Smaller related proposals:
fluxcd/pkg#1380 (kstatus message in wait errors) and fluxcd/notification-controller#1436
(message in the commit status).
All reactions