[SPARK-7831][Mesos] Added flag to shutdown driver when mesos dispatch…

[nraychaudhuri]

Fix for SPARK-7831

[nraychaudhuri]

@dragos@tnachen@skyluc Could you please take a look at this one?

[skyluc]

Instead of killing without failover, we could also start it without failover.

In the start method, to use:

valdriver= createSchedulerDriver(
master,
MesosClusterScheduler.this,
Utils.getCurrentUserName(),
appName,
conf,
Some(frameworkUrl),
Some(driverFailOver), // <-- with or without checkpoint dataSome(if (driverFailOver) Double.MaxValueelse0.0), // <-- timeout for failover recovery
fwId)

[nraychaudhuri]

Great find @skyluc

I will make the change

[tnachen]

jenkins please test

[tnachen]

Besides what @skyluc and my comments I think this patch LGTM. Have you tested this btw?

[nraychaudhuri]

Yes. I have tested this and it seems to work. I will make the necessary changes

[dragos]

ok to test

[dragos]

I confirm that the framework deregisters from Mesos. However, I don't see the old behavior anymore, where the framework stays even after stopping it. The new flag seems to have no effect.

$ sbin/start-mesos-dispatcher.sh --master mesos://lausanne1.local:5050
starting org.apache.spark.deploy.mesos.MesosClusterDispatcher, logging to /Users/dragos/workspace/Spark/dev/spark/logs/spark-dragos-org.apache.spark.deploy.mesos.MesosClusterDispatcher-1-sagitarius.local.out
$ sbin/stop-mesos-dispatcher.sh stopping org.apache.spark.deploy.mesos.MesosClusterDispatcher

The framework is gone.

[SparkQA]

Test build #50096 has finished for PR 10701 at commit 9002258.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[tnachen]

@dragos you mean the framework no longer shows up in the UI? the console output doesn't seem to suggest it's gone.

[dragos]

On 14 feb. 2016, at 10:01, Timothy Chen notifications@github.com wrote:

@dragos you mean the framework no longer shows up in the UI? the console output doesn't seem to suggest it's gone.

Yes, that's what I mean.

—
Reply to this email directly or view it on GitHub.

[tnachen]

I've tested this myself and is indeed now doing the correct behavior when not adding the flag in. I'll need to dig more, @nraychaudhuri have you tried this as well?

[tnachen]

I just found out that this is actually a bug in Mesos, where we cannot store a duration that's larger than int64_t. I filed a Mesos jira for this (https://issues.apache.org/jira/browse/MESOS-4862).
As a workaround, please don't use Double.MAX_VALUE but use Integer.MAX_VALUE instead which is what I did before, I forgot about hitting this in the past. We should also leave a comment to make sure we don't change this until it's fixed.

[andrewor14]

OK, let's not add a flag if it's a bug in Mesos. In the mean time before they fix it downstream we can use the workaround @tnachen suggested.

[dragos]

Sounds good. Who can close this PR?

[andrewor14]

@nraychaudhuri can you close this PR?

[tnachen]

@andrewor14@nraychaudhuri@dragos Sorry I'm not suggesting we close this PR, we still need the flag since we want to be able to either failover automatically or not. We only need to revert the particular line of change where the PR changed the timeout to DOUBLE.MAX_VALUE

[srowen]

@nraychaudhuri can you update or close this PR then?

[tnachen]

Seems like @nraychaudhuri is busy, I'll take this PR and update it myself. We definitely need this to be merged as it's quite useful for testing.