Ah that's right. I didn't mention it: I have 10 executors in my cluster, and so when I do .coalesce(10) and right after that saving orc to s3 - does coalescing really affects parallelism? To me it looks like no, because we went from 100 tasks that are executed in parallel by 10 executors to 10 tasks that are executed by same 10 executors. Now, I understand that there may be some data skew that may result in uneven partitions but that's not my case really (according to Spark UI). Again I'm trying to understand first of all how coalescing dataset impacts executor memory, gc etc. Maybe if coalesce is done before writing dataset, each of the resulting partition needs to be evaluated and thus stored in memory? - just a guess.
Andrii 2017-05-23 23:42 GMT+03:00 John Compitello <jo...@broadinstitute.org>: > Spark is doing operations on each partition in parallel. If you decrease > number of partitions, you’re potentially doing less work in parallel > depending on your cluster setup. > > On May 23, 2017, at 4:23 PM, Andrii Biletskyi <andrii.bilets...@yahoo.com. > INVALID <andrii.bilets...@yahoo.com.invalid>> wrote: > > > No, I didn't try to use repartition, how exactly it impacts the > parallelism? > In my understanding coalesce simply "unions" multiple partitions located > on same executor "one on on top of the other", while repartition does > hash-based shuffle decreasing the number of output partitions. So how this > exactly affects the parallelism, which stage of the job? > > Thanks, > Andrii > > > > On Tuesday, May 23, 2017 10:20 PM, Michael Armbrust < > mich...@databricks.com> wrote: > > > coalesce is nice because it does not shuffle, but the consequence of > avoiding a shuffle is it will also reduce parallelism of the preceding > computation. Have you tried using repartition instead? > > On Tue, May 23, 2017 at 12:14 PM, Andrii Biletskyi < > andrii.bilets...@yahoo.com.invalid> wrote: > > Hi all, > > I'm trying to understand the impact of coalesce operation on spark job > performance. > > As a side note: were are using emrfs (i.e. aws s3) as source and a target > for the job. > > Omitting unnecessary details job can be explained as: join 200M records > Dataframe stored in orc format on emrfs with another 200M records cached > Dataframe, the result of the join put back to emrfs. First DF is a set of > wide rows (Spark UI shows 300 GB) and the second is relatively small (Spark > shows 20 GB). > > I have enough resources in my cluster to perform the job but I don't like > the fact that output datasource contains 200 part orc files (as > spark.sql.shuffle. > partitions defaults to 200) so before saving orc to emrfs I'm doing > .coalesce(10). From documentation coalesce looks like a quite harmless > operations: no repartitioning etc. > > But with such setup my job fails to write dataset on the last stage. Right > now the error is OOM: GC overhead. When I change .coalesce(10) to > .coalesce(100) the job runs much faster and finishes without errors. > > So what's the impact of .coalesce in this case? And how to do in place > concatenation of files (not involving hive) to end up with smaller amount > of bigger files, as with .coalesce(100) job generates 100 orc snappy > encoded files ~300MB each. > > Thanks, > Andrii > > > > > >
