mirror of
https://github.com/k3s-io/kubernetes.git
synced 2026-01-19 08:15:01 +00:00
1be74e39fa
Automatic merge from submit-queue [Kubelet] Delay deletion of pod from the API server until volumes are deleted Previous PR that was reverted: #40239. To summarize the conclusion of the previous PR after reverting: - The status manager has the most up-to-date status, but the volume manager uses the status from the pod manager, which only is as up-to-date as the API server. - Because of this, the previous change required an additional round trip between the kubelet and API server. - When few pods are being added or deleted, this is only a minor issue. However, when under heavy load, the QPS limit to the API server causes this round trip to take ~60 seconds, which is an unacceptable increase in latency. Take a look at the graphs in #40239 to see the effect of QPS changes on timing. - To remedy this, the volume manager looks at the status from the status manager, which eliminates the round trip. cc: @vishh @derekwaynecarr @sjenning @jingxu97 @kubernetes/sig-storage-misc