NodeFeature resource is namespaced - undefined behaviour if same node in multiple namespaces? about node-feature-discovery HOT 4 CLOSED

Thanks @nobbs for the report. This is a recognized issue. There are measures we've taken to mitigate the problem of stale objects and labels. Which version of NFD were you using?

1. #1491 (released in NFD v0.15) added owner references to NodeFeature objects so that stale objects are GC'd (by Kubernetes) when the nfd worker pod that created it is deleted.
2. #1532 (will be released in NFD v0.16) is somewhat related (albeit does not help with stale NF objects) added a post-delete hook in the Helm chart to remove all NFD-generated labels when NFD gets uninstalled.
3. #1592 is an upcoming feature to limit/select the k8s namespaces where to watch NodeFeature objects.

For kustomize-based deployments doing kubectl delete -k ... deletes the nfd namespace and the stale NodeFeature objects as well. It will not cleanup the labels, though.

from node-feature-discovery.

Hey @marquiz - thanks for the info.

As we tried to track the releases nfd quite closely - most of the time delayed by 2-4 weeks - and the timeline the issue started, I'm pretty sure we migrated from one namespace to another either on one of the last v0.13.x patch releases or even on v0.14.0. The obsolete NodeFeature resources definitely had no owner references set at the time I've deleted them.

Based on the PRs you've linked it looks like the chance of running into this issue is basically naught with the recent / upcoming versions. Free to close this issue as resolved. :)

from node-feature-discovery.

@nobbs thx for reporting back. I'll close this for now but please reopen if you encounter any further issues.
/close

from node-feature-discovery.

@marquiz: Closing this issue.

from node-feature-discovery.