[
https://issues.apache.org/jira/browse/LUCENE-5399?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13870886#comment-13870886
]
Robert Muir commented on LUCENE-5399:
-------------------------------------
oh this is great, this was really frustrating when trying to test the missing
value support in DV.
as far as "my solution", i think its totally bogus actually: its caching a
reference to a mutable thing (BytesRef)... I just did it this way because
exactly one and only one thing calls this method (PagingFIeldCollector) and so
I knew i could do benchmarks and so on (i only investigated this as some 'page
2' results with searchAfter were like 8x slower than if you didnt use
searchAfter, with current patch like 2x).
as far as reversing the checks, that sounds fantastic, I think that explains
why my page 2 results are still 2x slower: 2x the work because they read the
ordinal twice, since most hits are neither competitive nor previously visited.
> PagingFieldCollector is very slow with String fields
> ----------------------------------------------------
>
> Key: LUCENE-5399
> URL: https://issues.apache.org/jira/browse/LUCENE-5399
> Project: Lucene - Core
> Issue Type: Bug
> Components: core/search
> Reporter: Robert Muir
> Attachments: LUCENE-5399.patch, LUCENE-5399.patch
>
>
> PagingFieldCollector (sort comparator) is significantly slower with string
> fields, because of how its "seen on a previous page" works: it calls
> compareDocToValue(int doc, T t) first to check this. (its the only user of
> this method)
> This is very slow with String, because no ordinals are used. so each document
> must lookup ord, then lookup bytes, then compare bytes.
> I think maybe we should replace this method with an 'after' slot, and just
> have compareDocToAfter or something.
> Otherwise we could use a hack-patch like the one i will upload (i did this
> just to test the performance, although tests do pass).
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
