Github user mccheah commented on the issue:
https://github.com/apache/spark/pull/21366
I think there's a single case here that could be missed.
If we delete a pod from the API server entirely, and neither the watch
receives a Pod with the deletion timestamp set OR the polling captures the pod
with a deletion timestamp set, and the pod just disappears from the cluster and
a subsequent poll doesn't return the pod at all - in that specific case we
won't mark the pod as deleted.
I think the above case shows that we're not being 100% accurate in a level
triggered approach. If you look carefully, we're only processing pod state as
seen by a periodic poll or watch events. But that isn't equivalent to
monitoring the whole cluster's state, periodically.
@foxish @dvogelbacher can you double check my theory here?
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
