[
https://issues.apache.org/jira/browse/CALCITE-3178?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16879907#comment-16879907
]
Gian Merlino commented on CALCITE-3178:
---------------------------------------
What's the best way to add a test case for a performance thing like that?
Should I add a case that checks for the correct result with a relatively short
timeout (a few seconds), so it's expected to fail in master, but potentially
succeed if the algorithm is improved? Any suggestion for the test class that
the test case should go into?
> RexSimplify.simplifyOrTerms slow with large OR filters
> ------------------------------------------------------
>
> Key: CALCITE-3178
> URL: https://issues.apache.org/jira/browse/CALCITE-3178
> Project: Calcite
> Issue Type: Improvement
> Affects Versions: 1.19.0
> Reporter: Gian Merlino
> Priority: Major
>
> In particular, once for each subpredicate within the OR,
> RexSimplify.simplifyOrTerms calls {{simplify.predicates.union}} and adds the
> freshly-unioned result to {{simplify.predicates}}. The most time-consuming
> part of this seems to be {{RexUtil.predicateConstants}}, which re-examines
> each previously-added entry. This is O(N^2) in the number of subpredicates
> within the OR.
> I discovered this when someone tried to run a query with a 14,000-element IN
> filter, and planning took about 45 seconds. In Druid, we always convert INs
> to ORs, never allowing Calcite's subquery conversion to happen. This is
> because as far as native Druid queries are concerned, a huge OR is going to
> be more efficient than a join against a constant subquery.
> I'm not sure what the best way is to fix this. The only thing that comes to
> mind immediately is the "quick fix" of limiting how many OR elements
> RexSimplify might attempt to simplify at once (and potentially AND as well? I
> haven't looked into that one.)
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
