[SPARK-2567] Resubmitted stage sometimes remains as active stage in the web UI

[tsudukim]

Moved the line which post SparkListenerStageSubmitted to the back of check of tasks size and serializability.

[AmplabJenkins]

Can one of the admins verify this patch?

[rxin]

How does this affect the final UI ?

[tsudukim]

You can see the screenshot which the original code generated in the JIRA.
https://issues.apache.org/jira/browse/SPARK-2567
This screenshot was taken after the job completed but one stage remained as Active Stage forever.
It shouldn't be displayed in this web ui at all because the corresponding new TaskSet is not submitted and even stage.newAttemptId() isn't called.
Sometimes this ghost stage appears when stage is re-submitted so this PR modified to prevent web ui from showing it.

[markhamstra]

This appears to be a reversion of d58502a while ignoring and misapplying the comment regarding ordering (which I'm not completely understanding.)

@xiajunluan ?

[tsudukim]

Hmm... I didn't notice it.
I'm going to rerun the test for confirmation as @xiajunluan 's commit comment.

[rxin]

Actually I aim to fix this in #1545

[tsudukim]

The test totally succeeded again.
If the @xiajunluan 's commit only aimed to avoid the unit test error, I think it should be reversioned as this PR. But I'm wondering if there were another aim.
@xiajunluan could you remember it?

[tsudukim]

Hi @rxin, thank you for following this ticket but couldn't we separate those problems into different PRs? SPARK-2298 is not about this problem.
I think we will be hard to trace why the code was modified and what discussion was made on the topic later. (just like we are now wondering about the intention of the commit of the last year)

[rxin]

Actually the fix looks good.

[rxin]

Jenkins, test this please.

[SparkQA]

QA tests have started for PR 1516. This patch merges cleanly.
View progress: https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/17063/consoleFull

[SparkQA]

QA results for PR 1516:
- This patch PASSES unit tests.
- This patch merges cleanly
- This patch adds no public classes

For more information see test ouptut:
https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/17063/consoleFull

[kayousterhout]

I looked into this a bit more and this change doesn't quite fix the problem in the right way, partially because of what @markhamstra pointed out (that we shouldn't send a SparkListenerStageCompleted event before sending a corresponding SparkListenerStageSubmitted event, which is I think what the mysterious comment was getting at) and partially because the bigger underlying problem here is that runningStages isn't updated at the right time (which also leads to a memory leak). I submitted an alternate fix here: #1566 -- let me know what you all think.

[tsudukim]

SPARK-2567 is resolved by #1566.