[
https://issues.apache.org/jira/browse/LUCENE-10574?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17538183#comment-17538183
]
Adrien Grand commented on LUCENE-10574:
---------------------------------------
I was assuming we wanted to have strong guarantees about the number of segments
in the index at search time, but it's a fair point that degrading to O(n^2)
merging to meet this guarantee is not a good trade-off.
I tried to think of ways we could do this. One obvious option is to remove
{{floorSegmentBytes}}, but this might be a bit too extreme as it would allow
any index to have a long tail of small segments? One idea I started playing
with consists of ensuring that every merge grows the largest input segment by
at least some fraction, e.g. 50%. It tries to strike a balance between avoiding
pathological merging and still trying to keep the number of segments contained
at search time. I quickly hacked this into TieredMergePolicy and this made the
StoredFieldsBenchmark more than 2x faster. I wonder if there are other
approaches we should consider.
> Remove O(n^2) from TieredMergePolicy or change defaults to one that doesn't
> do this
> -----------------------------------------------------------------------------------
>
> Key: LUCENE-10574
> URL: https://issues.apache.org/jira/browse/LUCENE-10574
> Project: Lucene - Core
> Issue Type: Bug
> Reporter: Robert Muir
> Priority: Major
>
> Remove {{floorSegmentBytes}} parameter, or change lucene's default to a merge
> policy that doesn't merge in an O(n^2) way.
> I have the feeling it might have to be the latter, as folks seem really wed
> to this crazy O(n^2) behavior.
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]