[
https://issues.apache.org/jira/browse/STORM-2809?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16249660#comment-16249660
]
Robert Joseph Evans commented on STORM-2809:
--------------------------------------------
So the supervisor was in waiting for blob localization on the slots it needed.
{code}
2017-11-10 16:39:36.629 o.a.s.d.s.Slot SLOT_6701 [INFO] STATE EMPTY msInState:
181025 -> WAITING_FOR_BLOB_LOCALIZATION msInState: 0
2017-11-10 16:39:36.629 o.a.s.d.s.Slot SLOT_6702 [INFO] STATE EMPTY msInState:
181026 -> WAITING_FOR_BLOB_LOCALIZATION msInState: 0
2017-11-10 16:39:36.633 o.a.s.d.s.Slot SLOT_6700 [INFO] STATE EMPTY msInState:
53 -> WAITING_FOR_BLOB_LOCALIZATION msInState: 0
{code}
Looks like 6700 was the bottleneck (as it was only in that state for 53 ms, not
the 3 mins of the others).
It was waiting for 3 mins to kill another worker.
{code}
2017-11-10 16:39:36.580 o.a.s.d.s.Slot SLOT_6700 [INFO] STATE KILL msInState:
179992 topo:TumblingWindowTestt500-18-1510353319 worker:null -> EMPTY
msInState: 0
{code}
Which looks a bit like turtles all they way down. It appears that the 3 min
timeout for both is causing them to just miss each other. Where the supervisor
gets in a bad state that takes 3 mins to recover from, but then the test times
out after 3 mins and kills the topology which causes the next topology to wait
3 mins to be launched....
I'll see if I can find the first one with issues.
> Integration test is failing consistently and topologies sometimes fail to
> start workers
> ---------------------------------------------------------------------------------------
>
> Key: STORM-2809
> URL: https://issues.apache.org/jira/browse/STORM-2809
> Project: Apache Storm
> Issue Type: Bug
> Affects Versions: 2.0.0
> Reporter: Stig Rohde Døssing
> Assignee: Robert Joseph Evans
> Fix For: 2.0.0
>
>
> The integration test has been failing fairly consistently since
> https://github.com/apache/storm/pull/2363. I tried running the test outside a
> VM with a locally installed Storm setup, and it has failed every time for me.
> Most runs seem to fail in ways that make it look like the integration test is
> just flaky (e.g. tuple windows not matching the calculated window), but in at
> least a few tests I saw the topology get submitted to Nimbus followed by
> about 3 minutes of nothing happening. The workers never started and the
> supervisor didn't seem aware of the scheduling. The only evidence that the
> topology was submitted was in the Nimbus log. This still happens even if the
> test topologies are killed with a timeout of 0, so there should be slots free
> for the next test immediately.
> I tried reverting https://github.com/apache/storm/pull/2363 and it seems to
> make the integration test pass much more often. Over 5 runs there was still
> an instance of a supervisor failing to start the workers, but the other 4
> passed.
> We should try to fix whatever is causing the supervisor to fail to start
> workers, and get the integration test more stable.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
