You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I suggest to drop the fsnotify based run-time reconfiguration of nfd-worker and nfd-master. All deployment methods (kustomize, Helm, operator) support (hash-based) pod restarts on configmap changes so we should use that. There have been issues reported on the current behavior (e.g. #805) where in some scenarios kubelet touches the mounted configmap every minute, causing a reconfiguration to happen every-minute. This behavior can be particularly detrimental on big clusters (of thousands of nodes) as reconfiguration of the nfd-master effectively restarts the object informers, causing it to resync object caches, putting heavy load on the apiserver/etcd.
Why is this needed:
Simplifies the code and reduces possible error (corner) scenarios
Simplifies the usage as there are never going to be config file settings that cannot be dynamically updated
Consistent functionality accross nfd-master, nfd-worker and nfd-topology-updater (topology-updater doesn't support dynamic runt-time reconfiguration)
The text was updated successfully, but these errors were encountered:
What would you like to be added:
I suggest to drop the fsnotify based run-time reconfiguration of nfd-worker and nfd-master. All deployment methods (kustomize, Helm, operator) support (hash-based) pod restarts on configmap changes so we should use that. There have been issues reported on the current behavior (e.g. #805) where in some scenarios kubelet touches the mounted configmap every minute, causing a reconfiguration to happen every-minute. This behavior can be particularly detrimental on big clusters (of thousands of nodes) as reconfiguration of the nfd-master effectively restarts the object informers, causing it to resync object caches, putting heavy load on the apiserver/etcd.
Why is this needed:
The text was updated successfully, but these errors were encountered: