I ran into this problem again, and I found that neither session.stop or newSession got rid of the error. So it's still not clear how to recover from a crashed(?) Spark session.

On the other hand, I did figure out why my sessions were crashing in the first place, so hopefully recovering from that will be a rare need. The reason is that wmfdata doesn't modify the default Spark when it starts a new session, which was (for example) causing it to start executors with only ~400 MiB of memory each.

I'm definitely going to change that, but it's not completely clear what the recommended settings for our cluster are. I cataloged the different recommendations at https://phabricator.wikimedia.org/T245097, and it would very helpful if one of y'all could give some clear recommendations about what the settings should be for local SWAP, YARN, and "large" YARN jobs. For example, is it important to increase spark.sql.shuffle.partitions for YARN jobs? Is it reasonable to use 8 GiB of driver memory for a local job when the SWAP servers only have 64 GiB total?

Thank you!

On Fri, 7 Feb 2020 at 06:53, Andrew Otto <otto@wikimedia.org> wrote:

Hm, interesting!  I don't think many of us have used SparkSession.builder.getOrCreate repeatedly in the same process.  What happens if you manually stop the spark session first, (session.stop()?) or maybe try to explicitly create a new session via newSession()?

On Thu, Feb 6, 2020 at 7:31 PM Neil Shah-Quinn <nshahquinn@wikimedia.org> wrote:

Hi Luca!

Those were separate Yarn jobs I started later. When I got this error, I found that the Yarn job corresponding to the SparkContext was marked as "successful", but I still couldn't get SparkSession.builder.getOrCreate to open a new one.

Any idea what might have caused that or how I could recover without restarting the notebook, which could mean losing a lot of in-progress work? I had already restarted that kernel so I don't know if I'll encounter this problem again. If I do, I'll file a task.

On Wed, 5 Feb 2020 at 23:24, Luca Toscano <ltoscano@wikimedia.org> wrote:

Hey Neil,

there were two Yarn jobs running related to your notebooks, I just killed them, let's see if it solves the problem (you might need to restart again your notebook). If not, let's open a task and investigate :)

Luca

Il giorno gio 6 feb 2020 alle ore 02:08 Neil Shah-Quinn <nshahquinn@wikimedia.org> ha scritto:

Whoa—I just got the same stopped SparkContext error on the query even after restarting the notebook, without an intermediate Java heap space error. That seems very strange to me.

On Wed, 5 Feb 2020 at 16:09, Neil Shah-Quinn <nshahquinn@wikimedia.org> wrote:

Hey there!

I was running SQL queries via PySpark (using the wmfdata package) on SWAP when one of my queries failed with "java.lang.OutofMemoryError: Java heap space".

After that, when I tried to call the spark.sql function again (via wmfdata.hive.run), it failed with "java.lang.IllegalStateException: Cannot call methods on a stopped SparkContext."

When I tried to create a new Spark context using SparkSession.builder.getOrCreate (whether using wmfdata.spark.get_session or directly), it returned a SparkContent object properly, but calling the object's sql function still gave the "stopped SparkContext error".

Any idea what's going on? I assume restarting the notebook kernel would take care of the problem, but it seems like there has to be a better way to recover.

Thank you!

_______________________________________________
Analytics mailing list
Analytics@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics

_______________________________________________
Analytics mailing list
Analytics@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics

_______________________________________________
Analytics mailing list
Analytics@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics

_______________________________________________
Analytics mailing list
Analytics@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics