[
https://issues.apache.org/jira/browse/LUCENE-5675?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14009545#comment-14009545
]
Michael McCandless commented on LUCENE-5675:
--------------------------------------------
Thanks Steve!
> "ID postings format"
> --------------------
>
> Key: LUCENE-5675
> URL: https://issues.apache.org/jira/browse/LUCENE-5675
> Project: Lucene - Core
> Issue Type: New Feature
> Affects Versions: 4.9, 5.0
> Reporter: Robert Muir
> Fix For: 4.9, 5.0
>
> Attachments: LUCENE-5675.patch
>
>
> Today the primary key lookup in lucene is not that great for systems like
> solr and elasticsearch that have versioning in front of IndexWriter.
> To some extend BlockTree can "sometimes" help avoid seeks by telling you the
> term does not exist for a segment. But this technique (based on FST prefix)
> is fragile. The only other choice today is bloom filters, which use up huge
> amounts of memory.
> I don't think we are using everything we know: particularly the version
> semantics.
> Instead, if the FST for the terms index used an algebra that represents the
> max version for any subtree, we might be able to answer that there is no term
> T with version < V in that segment very efficiently.
> Also ID fields dont need postings lists, they dont need stats like
> docfreq/totaltermfreq, etc this stuff is all implicit.
> As far as API, i think for users to provide "IDs with versions" to such a PF,
> a start would to set a payload or whatever on the term field to get it thru
> indexwriter to the codec. And a "consumer" of the codec can just cast the
> Terms to a subclass that exposes the FST to do this version check efficiently.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
