[
https://issues.apache.org/jira/browse/SOLR-5344?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15629832#comment-15629832
]
James Dyer commented on SOLR-5344:
----------------------------------
ok, I think I know what's going here. This feature is supposed to estimate hit
counts for spelling corrections for cases where the client doesn't care about
the exact # of hits, only that a partcular collation, if re-queried, would
return something. To gets estimates, you tell it the max # of documents you
would like it to collect before quitting. It then estimates how many hits it
would have counted with this:
{noformat}
maximum-doc-id * number-of-docs-collected / (# visited docs + last-doc-id + 1)
{noformat}
In the failing test, we ask it to collect between 5 and 20 documents. The
max-doc-id is always 17 (there are 17 documents and no deletions).
But the denominator is controlled by the # of visited documents, and also the
doc id of the one that happened to be visited last. But in the face of
randomized testing and release-specific index behavior, I think the best we can
hope for is a worse-case scenario, between 2 and 15. The actual correct value
is 8.
So unless there are objections, I am going to relax the requirement of 6 <=
hits <= 10 , and use 2 <= hits <= 15. Maybe we could do better than this, but
I would think anyone using this feature probably does not need to know more
than whether or not hits can be produced, or the relative # between several
collations returned.
> SpellCheckCollatorTest.testEstimatedHitCounts fails in jenkins from time to
> time
> --------------------------------------------------------------------------------
>
> Key: SOLR-5344
> URL: https://issues.apache.org/jira/browse/SOLR-5344
> Project: Solr
> Issue Type: Bug
> Reporter: Robert Muir
> Assignee: James Dyer
>
> Doesn't happen very often, but maybe one I can fix. I'll look into it.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
