lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Michael McCandless (JIRA)" <>
Subject [jira] Commented: (LUCENE-2271) Function queries producing scores of -inf or NaN (e.g. 1/x) return incorrect results with TopScoreDocCollector
Date Sun, 21 Feb 2010 15:37:27 GMT


Michael McCandless commented on LUCENE-2271:

This is a delightfully clever solution, delegating to a contained
collector and then swapping in one collector for the startup (when the
queue is not yet full) and then a new collector once the queue is

It's the closest equivalent we in javaland can reach, to using
function pointers in C.

But, I don't think we should commit this -- this makes the code more
complex, and doesn't really gain enough (performance is the same) in
return.  It also relies more heavily on the JVM being able to optimize away
the method call.

Yes it handles -inf/nan, but I don't think Lucene's default
sort-by-relevance collector needs to (prior to 2.9 we silently
filtered out such hits, as well as all hits with score <= 0.0).

> Function queries producing scores of -inf or NaN (e.g. 1/x) return incorrect results
with TopScoreDocCollector
> --------------------------------------------------------------------------------------------------------------
>                 Key: LUCENE-2271
>                 URL:
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: Search
>    Affects Versions: 2.9
>            Reporter: Uwe Schindler
>            Priority: Minor
>             Fix For: 2.9.2, 3.0.1, 3.1
>         Attachments: LUCENE-2271-bench.patch, LUCENE-2271.patch, LUCENE-2271.patch, LUCENE-2271.patch,
LUCENE-2271.patch, TSDC.patch
> This is a foolowup to LUCENE-2270, where a part of this problem was fixed (boost = 0
leading to NaN scores, which is also un-intuitive), but in general, function queries in Solr
can create these invalid scores easily. In previous version of Lucene these scores ordered
correct (except NaN, which mixes up results), but never invalid document ids are returned
(like Integer.MAX_VALUE).
> The problem is: TopScoreDocCollector pre-fills the HitQueue with sentinel ScoreDocs with
a score of -inf and a doc id of Integer.MAX_VALUE. For the HQ to work, this sentinel must
be smaller than all posible values, which is not the case:
> - -inf is equal and the document is not inserted into the HQ, as not competitive, but
the HQ is not yet full, so the sentinel values keep in the HQ and result is the Integer.MAX_VALUE
docs. This problem is solveable (and only affects the Ordered collector) by chaning the exit
condition to:
> {code}
> if (score <= pqTop.score && pqTop.doc != Integer.MAX_VALUE) {
>     // Since docs are returned in-order (i.e., increasing doc Id), a document
>     // with equal score to pqTop.score cannot compete since HitQueue favors
>     // documents with lower doc Ids. Therefore reject those docs too.
>     return;
> }
> {code}
> - The NaN case can be fixed in the same way, but then has another problem: all comparisons
with NaN result in false (none of these is true): x < NaN, x > NaN, NaN == NaN. This
leads to the fact that HQ's lessThan always returns false, leading to unexspected ordering
in the PQ and sometimes the sentinel values do not stay at the top of the queue. A later hit
then overrides the top of the queue but leaves the incorrect sentinels  unchanged -> invalid
results. This can be fixed in two ways in HQ:
> Force all sentinels to the top:
> {code}
> protected final boolean lessThan(ScoreDoc hitA, ScoreDoc hitB) {
>     if (hitA.doc == Integer.MAX_VALUE)
>       return true;
>     if (hitB.doc == Integer.MAX_VALUE)
>       return false;
>     if (hitA.score == hitB.score)
>       return hitA.doc > hitB.doc; 
>     else
>       return hitA.score < hitB.score;
> }
> {code}
> or alternatively have a defined order for NaN ( sorts them after +inf):
> {code}
> protected final boolean lessThan(ScoreDoc hitA, ScoreDoc hitB) {
>     if (hitA.score == hitB.score)
>       return hitA.doc > hitB.doc; 
>     else
>       return, hitB.score) < 0;
> }
> {code}
> The problem with both solutions is, that we have now more comparisons per hit and the
use of sentinels is questionable. I would like to remove the sentinels and use the old pre
2.9 code for comparing and using PQ.add() when a competitive hit arrives. The order of NaN
would be unspecified.
> To fix the order of NaN, it would be better to replace all score comparisons by
[also in FieldComparator].
> I would like to delay 2.9.2 and 3.0.1 until this problem is discussed and solved.

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message