we have found that to make shuffles reliable without OOMs we need to have spark.sql.shuffle.partitions at a high number, bigger than 2000 at least. yet this leads to a large amount of part files, which puts big pressure on spark driver programs.
i tried to mitigate this with dataframe.coalesce to reduce the number of files, but this is not acceptable. coalesce changes the tasks for the last shuffle before it, bringing back the issues we tried to mitigate with a high number for spark.sql.shuffle.partitions in the first place. doing a dataframe.repartition before every write is also not an unacceptable approach, it is too high a price to pay just to bring down the number of files. so i am very excited about any approach that efficiently merges files when writing. On Mon, Aug 6, 2018 at 5:28 PM, lukas nalezenec <lu...@apache.org> wrote: > Hi Koert, > There is no such Jira yet. We need SPARK-23889 before. You can find some > mentions in the design document inside 23889. > Best regards > Lukas > > 2018-08-06 18:34 GMT+02:00 Koert Kuipers <ko...@tresata.com>: > >> i went through the jiras targeting 2.4.0 trying to find a feature where >> spark would coalesce/repartition by size (so merge small files >> automatically), but didn't find it. >> can someone point me to it? >> thank you. >> best, >> koert >> >> On Sun, Aug 5, 2018 at 9:06 PM, Koert Kuipers <ko...@tresata.com> wrote: >> >>> lukas, >>> what is the jira ticket for this? i would like to follow it's activity. >>> thanks! >>> koert >>> >>> On Wed, Jul 25, 2018 at 5:32 PM, lukas nalezenec <lu...@apache.org> >>> wrote: >>> >>>> Hi, >>>> Yes, This feature is planned - Spark should be soon able to repartition >>>> output by size. >>>> Lukas >>>> >>>> >>>> Dne st 25. 7. 2018 23:26 uživatel Forest Fang <forest.f...@outlook.com> >>>> napsal: >>>> >>>>> Has there been any discussion to simply support Hive's merge small >>>>> files configuration? It simply adds one additional stage to inspect size >>>>> of >>>>> each output file, recompute the desired parallelism to reach a target >>>>> size, >>>>> and runs a map-only coalesce before committing the final files. Since >>>>> AFAIK >>>>> SparkSQL already stages the final output commit, it seems feasible to >>>>> respect this Hive config. >>>>> >>>>> https://community.hortonworks.com/questions/106987/hive-mult >>>>> iple-small-files.html >>>>> >>>>> >>>>> On Wed, Jul 25, 2018 at 1:55 PM Mark Hamstra <m...@clearstorydata.com> >>>>> wrote: >>>>> >>>>>> See some of the related discussion under https://github.com/apach >>>>>> e/spark/pull/21589 >>>>>> >>>>>> If feels to me like we need some kind of user code mechanism to >>>>>> signal policy preferences to Spark. This could also include ways to >>>>>> signal >>>>>> scheduling policy, which could include things like scheduling pool and/or >>>>>> barrier scheduling. Some of those scheduling policies operate at >>>>>> inherently >>>>>> different levels currently -- e.g. scheduling pools at the Job level >>>>>> (really, the thread local level in the current implementation) and >>>>>> barrier >>>>>> scheduling at the Stage level -- so it is not completely obvious how to >>>>>> unify all of these policy options/preferences/mechanism, or whether it is >>>>>> possible, but I think it is worth considering such things at a fairly >>>>>> high >>>>>> level of abstraction and try to unify and simplify before making things >>>>>> more complex with multiple policy mechanisms. >>>>>> >>>>>> On Wed, Jul 25, 2018 at 1:37 PM Reynold Xin <r...@databricks.com> >>>>>> wrote: >>>>>> >>>>>>> Seems like a good idea in general. Do other systems have similar >>>>>>> concepts? In general it'd be easier if we can follow existing >>>>>>> convention if >>>>>>> there is any. >>>>>>> >>>>>>> >>>>>>> On Wed, Jul 25, 2018 at 11:50 AM John Zhuge <jzh...@apache.org> >>>>>>> wrote: >>>>>>> >>>>>>>> Hi all, >>>>>>>> >>>>>>>> Many Spark users in my company are asking for a way to control the >>>>>>>> number of output files in Spark SQL. There are use cases to either >>>>>>>> reduce >>>>>>>> or increase the number. The users prefer not to use function >>>>>>>> *repartition*(n) or *coalesce*(n, shuffle) that require them to >>>>>>>> write and deploy Scala/Java/Python code. >>>>>>>> >>>>>>>> Could we introduce a query hint for this purpose (similar to >>>>>>>> Broadcast Join Hints)? >>>>>>>> >>>>>>>> /*+ *COALESCE*(n, shuffle) */ >>>>>>>> >>>>>>>> In general, is query hint is the best way to bring DF functionality >>>>>>>> to SQL without extending SQL syntax? Any suggestion is highly >>>>>>>> appreciated. >>>>>>>> >>>>>>>> This requirement is not the same as SPARK-6221 that asked for >>>>>>>> auto-merging output files. >>>>>>>> >>>>>>>> Thanks, >>>>>>>> John Zhuge >>>>>>>> >>>>>>> >>> >> >
