[
https://issues.apache.org/jira/browse/SPARK-9066?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14662707#comment-14662707
]
Weizhong commented on SPARK-9066:
---------------------------------
Yes, the root reaason is same, that is cause by scan HDFS too many times, in
[PR#6454|https://github.com/apache/spark/pull/6454] use coalesce to decrease
partitions, but add two shuffles, but if we change the cartesian order also can
decrease the scan times, which I have done in
[PR#7417|https://github.com/apache/spark/pull/7417]
> Improve cartesian performance
> ------------------------------
>
> Key: SPARK-9066
> URL: https://issues.apache.org/jira/browse/SPARK-9066
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Reporter: Weizhong
> Priority: Minor
>
> Currently, for CartesianProduct, if right plan partition record number are
> small than left partition record number, then the performance is bad as need
> do many times scan for right plan.
> For example:
> {noformat}
> with single_value as (
> select max(1) tpcds_val from date_dim
> )
> select sum(ss_quantity * ss_sales_price) ssales, tpcds_val
> from store_sales, single_value
> group by tpcds_val
> {noformat}
> above SQL clause, right plan only have 1 record, left plan have 1823
> partiton(in our test) and each partition has more than 4000 records, then for
> each left plan partition record we need scan data from hdfs for right plan.
> That is, for left plan we need scan _left_plan_partition_num_ times, for
> right plan we need scan _left_plan_partition_num * right_plan_partition_num_
> times, total is _left_plan_partition_num * (1 + right_plan_partition_num)_
> times
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
