Dmitry Tikhonov created SOLR-18676:
--------------------------------------
Summary: Combined Query (RRF): fuse on a single core in standalone
and embedded Solr
Key: SOLR-18676
URL: https://issues.apache.org/jira/browse/SOLR-18676
Project: Solr
Issue Type: Improvement
Components: Embedded, search
Reporter: Dmitry Tikhonov
Combined Query (SOLR-17319) always forces distributed search:
CombinedQuerySearchHandler
calls setForcedDistrib(true), and the sub-query lists are fused only in
CombinedQueryComponent.mergeIds on the coordinator. Outside SolrCloud this has
three
consequences:
* Embedded Solr (EmbeddedSolrServer) cannot run a combined query at all: with
no shards
parameter, HttpShardHandler.prepDistributed builds the shard URL from
req.getHttpSolrCall(), which is null there.
* Standalone Solr sends every combined query to itself over HTTP, and needs its
own URL
in solr.security.allow.urls (documented as a requirement).
* With distrib=false, process() returns QueryAndResponseCombiner.simpleCombine:
a union
in HashMap order, with no RRF and no sort.
Proposal: outside SolrCloud, a combined query that names no shards (and does
not ask for
distrib=true) is fused on the core that receives it. The local path is mergeIds
over the
lists of a single shard, sharing its code (union and collapse across
sub-queries, the
configured combiner, the request's sort, numFound and numFoundExact, combiner
explanations). A single core therefore answers exactly what a one-shard
distributed
search over that core answers. Each sub-query is fetched to start + rows, or
shards.rows
when given, as a shard would be. SolrCloud behaviour is unchanged: the
coordinator still
always fuses.
While comparing the local and distributed answers, two existing bugs of the
distributed
path came up, fixed in the same change (they can be split out):
1. A single-pass distributed combined query (e.g. fl=id,score) returns one
sub-query's
raw score for each document instead of its fused (RRF) score; a two-pass
request
returns the fused score. CombinedQueryComponent now overrides returnFields
to set the
fused score.
2. Elevation from elevate.xml that matches the text of only some sub-queries
fails the
request with IndexOutOfBoundsException (SolrCloud included).
QueryElevationComponent
copies the first sub-query's elevated SortSpec onto the request, but each
sub-query's
sort values are computed for its own sort, which then lacks the _elevate_
field. Sort
values are now computed for the request's sort for every sub-query.
Behaviour change: standalone Solr no longer distributes a combined query to
itself, so
it no longer needs its own URL allow-listed, and distrib=false returns the
fused, sorted
result.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]