[
https://issues.apache.org/jira/browse/OAK-10648?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17817235#comment-17817235
]
Thomas Mueller commented on OAK-10648:
--------------------------------------
I didn't test this yet, but the following change seem to be necessary:
{noformat}
oak-search FulltextIndexPlanner
if (pr.isNotNullRestriction()) {
// don't use weight for "is not null" restrictions
weight = 1;
---------------- missing code start --------------------------
} else if (pr.isNullRestriction()) {
// don't use weight for "is null" restrictions
weight = 1;
---------------- missing code end --------------------------
} else {
if (weight > 1) {
// for non-equality conditions such as
// where x > 1, x < 2, x like y,...:
// use a maximum weight of 3,
// so assume we read at least 30%
if (!isEqualityRestriction(pr)) {
weight = Math.min(3, weight);
}
}
}
{noformat}
We should probably add a feature toggle / system property so that we can switch
back to the original behavior, to we can switch back in case an application
relies on the current behavior.
> Null Props Cause Incorrect Query Estimation
> -------------------------------------------
>
> Key: OAK-10648
> URL: https://issues.apache.org/jira/browse/OAK-10648
> Project: Jackrabbit Oak
> Issue Type: Bug
> Components: indexing
> Reporter: Patrique Legault
> Priority: Major
> Attachments: Non Union Query Plan.json, Non Union With Null
> Check.json, Screenshot 2024-02-13 at 9.30.43 AM.png, Union Query Plan.json,
> cqTagLucene.json
>
>
> Using null props in a query can cause the query engine to incorrectly
> estimate the cost of query plan which can lead to a traversal and slow
> queries to execute.
>
> If you look at the query plan below the number of null props documents is
> quiet high yet the cost for the query is only 19. When we execute the UNION
> query the cost is 38 which is why it is not selected when in reality the
> original cost should be much higher.
>
> After removing the null check the cost estimation is drastically different
> and correctly reflects the number of documents in the index.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
