Ian Springer created SOLR-18625:
-----------------------------------
Summary: eDisMax adds an empty pf clause when no phrase qualifies,
defeating the MatchAllDocsQuery optimization
Key: SOLR-18625
URL: https://issues.apache.org/jira/browse/SOLR-18625
Project: Solr
Issue Type: Bug
Components: query parsers
Affects Versions: 10.0, 9.9
Reporter: Ian Springer
{code}
When \{{pf}} is set and no phrase qualifies for the phrase boost, eDisMax adds
an empty \{{BooleanQuery}} to the main query as a \{{SHOULD}} clause. This
happens with single-term queries and with \{{q=:}}. It shows up as \{{()}} in
the parsed query, e.g.:
{noformat}
defType=edismax&qf=subject&pf=subject&q=wireless
-> +(subject:wireless) ()
{noformat}
Cause: in \{{ExtendedDismaxQParser.addShingledPhraseQueries}}, the pf
sub-parser runs with \{{minClauseSize=2}}. That drops every one-position
phrase, so \{{pp.parse(...)}} returns an empty \{{BooleanQuery}} instead of
\{{null}}. The \{{phrase != null}} guard therefore lets it through.
Impact:
- Scoring is unaffected, because an empty SHOULD clause matches nothing.
- It defeats the single-clause \{{MatchAllDocsQuery}} unwrapping in
\{{ExtendedDismaxQParser.parse()}}. With \{{pf}} set, \{{q=:}} therefore scores
a \{{BooleanQuery}} over every doc. It misses Solr's \{{MatchAllDocsQuery}}
fast paths (answered from the live-docs set, bypassing the filterCache).
- Fielded-only queries such as \{{q=subject:foo}} produce no pf clauses. They
still create a sub-parser and analyze an empty phrase for each pf slop group.
Proposed fix:
- Skip adding the phrase clause when the parsed result is empty, not just when
it is null.
- Return early from \{{addShingledPhraseQueries}} when there are no clauses.
Ranking does not change. This was originally part of SOLR-18314 /
https://github.com/apache/solr/pull/4652, which was closed because of its
separate DisMax ranking change. Reviewers agreed this part was safe to merge on
its own.
{code}
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
