Mailing-List: contact dev-help@lucene.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@lucene.apache.org
Date: Thu, 25 Jun 2015 01:48:06 +0000 (UTC)
From: "ASF subversion and git services (JIRA)" <jira@apache.org>
To: dev@lucene.apache.org
Message-ID: <JIRA.12838993.1434670951000.10555.1435196886485@Atlassian.JIRA>
In-Reply-To: <JIRA.12838993.1434670951000@Atlassian.JIRA>
References: <JIRA.12838993.1434670951000@Atlassian.JIRA>
 <JIRA.12838993.1434670951639@arcas>
Subject: [jira] [Commented] (LUCENE-6585) Make ConjunctionDISI flatten sub
 ConjunctionDISI instances
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit


    [ https://issues.apache.org/jira/browse/LUCENE-6585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14600523#comment-14600523 ] 

ASF subversion and git services commented on LUCENE-6585:
---------------------------------------------------------

Commit 1687406 from [~rcmuir] in branch 'dev/branches/branch_5x'
[ https://svn.apache.org/r1687406 ]

LUCENE-6585: add some tests that coord is applied properly for nested conjunctions

> Make ConjunctionDISI flatten sub ConjunctionDISI instances
> ----------------------------------------------------------
>
>                 Key: LUCENE-6585
>                 URL: https://issues.apache.org/jira/browse/LUCENE-6585
>             Project: Lucene - Core
>          Issue Type: Improvement
>            Reporter: Adrien Grand
>            Priority: Minor
>         Attachments: LUCENE-6585.patch
>
>
> Today ConjunctionDISI wraps some sub (two-phase) iterators. I would like to improve it by flattening sub iterators when they implement ConjunctionDISI. In practice, this would make "+A +(+B +C)" be executed more like "+A +B +C" (only in terms of matching, scoring would not change).
> My motivation for this is that if we don't flatten and are unlucky, we can sometimes hit some worst cases. For instance consider the 3 following postings lists (sorted by increasing cost):
> A: 1, 1001, 2001, 3001, ...
> C: 0, 2, 4, 6, 8, 10, 12, 14, ...
> B: 1, 3, 5, 7, 9, 11, 13, 15, ...
> If we run "+A +B +C", then everything works fine, we use A as a lead, and advance B 1000 by 1000 to find the next match (if any).
> However if we run "+A +(+B +C)", then we would iterate B and C 2 by 2 over the entire doc ID space when trying to find the first match which occurs on or after A:1.
> This is an extreme example which is unlikely to happen in practice, but flattening would also help a bit on some more common cases. For instance imagine that A, B and C have respective costs of 100, 10 and 1000. If you search for "+A +(+B +C)", then we will use the most costly iterator (C) to confirm matches of B (the least costly iterator, used as a lead) while it would have been more efficient to confirm matches of B with A first, since A is less costly than C.


--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org