[
https://issues.apache.org/jira/browse/SPARK-47836?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17836606#comment-17836606
]
Tanel Kiis commented on SPARK-47836:
------------------------------------
I would be willing to make a PR, but I do not know the right solution.
Migrating from a custom solution to the KllDoublesSketch would be the cleanest,
but its results are nondeterministic
> Performance problem with QuantileSummaries
> ------------------------------------------
>
> Key: SPARK-47836
> URL: https://issues.apache.org/jira/browse/SPARK-47836
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 3.5.1
> Reporter: Tanel Kiis
> Priority: Major
>
> SPARK-29336 caused a severe performance regression.
> In practice a partial_aggregate with several approx_percentile calls ran less
> than hour and the final aggrergation after exchange would have taken over a
> week.
> Simple percentile ran about the same time in the first part and the final
> aggregate ran very quickly.
> I made a benchmark, and it reveals that the merge operation is very-very
> slow:
> https://github.com/tanelk/spark/commit/3b16f429a77b10003572295f42361fbfb2f3c63e
> From my experiments it looks like it is n^2 with the number of partitions
> (number of partial aggregations to merge).
> When I reverted the changes made in this PR, then the "Only insert" and
> "Insert & merge" were very similar.
> The cause seems to be, that compressImmut does not reduce the number samples
> allmost at all after merges and just keeps iterating over an evergrowing list.
> I was not able to figure out how to fix the issue without just reverting the
> PR.
> I also added a benchmark with KllDoublesSketch from the apache datasketches
> project and it worked even better than this class before this PR.
> Only downside was that it is not-deterministic.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
