[
https://issues.apache.org/jira/browse/APEXCORE-595?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15793301#comment-15793301
]
Tushar Gosavi commented on APEXCORE-595:
----------------------------------------
yes, that is the scenario when all physical operators are removed then there
are no partitions for root operators and committedWindowId in
UpdateCheckpointsContext never gets updated from default value (Long.MAX_VALUE).
I observed this behavior in pi demo application with modified random generator
which raise ShutdownException after some windows. some containers failed trying
to purge Long.MAX_VALUE window from buffer server. Need to open another issue
for bufferserver behavior while trying to purge non existed window.
IMO the suitable fix is keep committedWindowId unchanged when there are no
physical operators present in DAG. let me know what do you think.
> Master incorrectly updates committedWindowId when all partitions are
> terminated.
> --------------------------------------------------------------------------------
>
> Key: APEXCORE-595
> URL: https://issues.apache.org/jira/browse/APEXCORE-595
> Project: Apache Apex Core
> Issue Type: Bug
> Reporter: Tushar Gosavi
> Assignee: Tushar Gosavi
>
> When all partitions are terminated application master incorrectly updates the
> committedWindowId to Long.MAX_VALUE and send purge requet to StramChild
> causing exception in StramChild.
> The problematic code is in StreamingContainerManager.updateCheckpoints.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
