Both cleanup rely on the periodic syncBatch in the status manger to trigger the final pod deletion. This is specifically checked in if m.needsUpdate(syncedUID, status) || m.couldBeDeleted(uid, status.status) .
However, syncBatch calls syncPod, which checks if !m.needsUpdate(uid, status), and skip if it's true at the beginning of the function. This means that it will never trigger the final deletion request (unless something else triggers it).
This causes pod deletion to take > 1 minute in many cases, hence a bunch of test flake.
Ideally, kubelet should not rely on the periodic syncBatch to send out deletion request, but at the least, this buggy logic should be fixed.
/cc @kubernetes/sig-node-bugs @dchen1107 @Random-Liu
Both cleanup rely on the periodic
syncBatchin the status manger to trigger the final pod deletion. This is specifically checked inif m.needsUpdate(syncedUID, status) || m.couldBeDeleted(uid, status.status).However,
syncBatchcallssyncPod, which checksif !m.needsUpdate(uid, status), and skip if it's true at the beginning of the function. This means that it will never trigger the final deletion request (unless something else triggers it).This causes pod deletion to take > 1 minute in many cases, hence a bunch of test flake.
Ideally, kubelet should not rely on the periodic
syncBatchto send out deletion request, but at the least, this buggy logic should be fixed./cc @kubernetes/sig-node-bugs @dchen1107 @Random-Liu