Atri Sharma created LUCENE-8824:
-----------------------------------
Summary: TestTopDocsMerge Is Broken
Key: LUCENE-8824
URL: https://issues.apache.org/jira/browse/LUCENE-8824
Project: Lucene - Core
Issue Type: Improvement
Reporter: Atri Sharma
Investigating a test failure post-LUCENE-8757, I realized that TestTopDocsMerge
takes a non-obvious invariant on the fact that the number of Collectors
involved in the merge will be equal to the number of LeafReaderContexts
originally present. This is propagated in the corresponding ScoreDocs's
shardIndex fields, which can lead to subtle issues since shardIndex is used for
tie-breaking in the priority queue used during the merge. I believe that this
is a dangerous and unnecessary dependency to take since the
IndexSearcher#slices method does not advertise any such guarantees.
The underlying assumption worked well in the past since the default slice
allocation algorithm always allocated a slice per segment, thus guaranteeing
that the number of Collectors (== number of Slices) will be equal to the number
of Leaf Contexts. With 8757, this is no longer true.
I propose a rewrite of the test, where ShardSearcher is allowed to take a
LeafSlice instance and can internally do a sequential search on the leaf
contexts of the passed in the slice. This will allow TestTopDocsMerge to create
N subsearchers where N is equal to the number of slices used by the
IndexSearcher being compared to.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
