spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Michael Armbrust <mich...@databricks.com>
Subject Re: Question about data frame partitioning in Spark 1.3.0
Date Wed, 14 Oct 2015 17:11:13 GMT
This won't help as for two reasons:
 1) Its all still just creating lineage since you aren't caching the
partitioned data.  It will still fetch the shuffled blocks for each query.
 2) The query optimizer is not aware of RDD level partitioning since its
mostly a blackbox.

1) could be fixed by adding caching.  2) is on our roadmap (though you'd
have to use logical DataFrame expressions to do the partitioning instead of
a class based partitioner).

On Wed, Oct 14, 2015 at 8:45 AM, Cesar Flores <cesar7@gmail.com> wrote:

>
> My current version of spark is 1.3.0 and my question is the next:
>
> I have large data frames where the main field is an user id. I need to do
> many group by's and joins using that field. Do the performance will
> increase if before doing any group by or join operation I first convert to
> rdd to partition by the user id? In other words trying something like the
> next lines in all my user data tables will improve the performance in the
> long run?:
>
> val partitioned_rdd = unpartitioned_df
>    .map(row=>(row.getLong(0), row))
>    .partitionBy(new HashPartitioner(200))
>    .map(x => x._2)
>
> val partitioned_df = hc.createDataFrame(partitioned_rdd,
> unpartitioned_df.schema)
>
>
>
>
> Thanks a lot
> --
> Cesar Flores
>

Mime
View raw message