[
https://issues.apache.org/jira/browse/LUCENE-5409?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13878749#comment-13878749
]
Michael McCandless commented on LUCENE-5409:
--------------------------------------------
Hi Peng, this is a good catch! Probably we could key on the .origChildQuery
since we already carry that one along.
Do you have a test case showing the issue?
> ToParentBlockJoinCollector.getTopGroups returns empty Groups
> ------------------------------------------------------------
>
> Key: LUCENE-5409
> URL: https://issues.apache.org/jira/browse/LUCENE-5409
> Project: Lucene - Core
> Issue Type: Bug
> Components: core/search
> Affects Versions: 4.6
> Environment: Ubuntu 12.04
> Reporter: Peng Cheng
> Assignee: Michael McCandless
> Priority: Critical
> Fix For: 4.7
>
> Original Estimate: 168h
> Remaining Estimate: 168h
>
> A bug is observed to cause unstable results returned by the getTopGroups
> function of class ToParentBlockJoinCollector.
> In the scorer generation stage, the ToParentBlockJoinCollector will
> automatically rewrite all the associated ToParentBlockJoinQuery (and their
> subqueries), and save them into its in-memory Look-up table, namely
> joinQueryID (see enroll() method for detail). Unfortunately, in the
> getTopGroups method, the new ToParentBlockJoinQuery parameter is not
> rewritten (at least users are not expected to do so). When the new one is
> searched in the old lookup table (considering the impact of rewrite() on
> hashCode()), the lookup will largely fail and eventually end up with a
> topGroup collection consisting of only empty groups (their hitCounts are
> guaranteed to be zero).
> An easy fix would be to rewrite the original BlockJoinQuery before invoking
> getTopGroups method. However, the computational cost of this is not optimal.
> A better but slightly more complex solution would be to save unrewrited
> Queries into the lookup table.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
