jpountz closed pull request #227: LUCENE-10033: Encode numeric doc values and
ordinals of SORTED(_SET) doc values in blocks.
URL: https://github.com/apache/lucene/pull/227
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use t
shubhamvishu commented on code in PR #11832:
URL: https://github.com/apache/lucene/pull/11832#discussion_r988712945
##
lucene/core/src/java/org/apache/lucene/index/SlowCodecReaderWrapper.java:
##
@@ -163,7 +163,7 @@ private static KnnVectorsReader
readerToVectorReader(LeafReade
rmuir commented on issue #11839:
URL: https://github.com/apache/lucene/issues/11839#issuecomment-1269611646
@zhaih please take it!
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific com
jpountz commented on PR #11832:
URL: https://github.com/apache/lucene/pull/11832#issuecomment-1269629285
> One reasoning is we can have this to be consistent like how we have
similar static functions in DocValues. There no code using it right now but
this could be used maybe in future?
jpountz commented on PR #11831:
URL: https://github.com/apache/lucene/pull/11831#issuecomment-1269633237
Why did we have to update the number of allowed deletes back to 33% on some
tests, did they fail otherwise? Is there another way how we could improve these
tests to cope better with the
jpountz commented on PR #11831:
URL: https://github.com/apache/lucene/pull/11831#issuecomment-1269635726
E.g. maybe it would be better to add more documents in
`testNRTIsCurrentAfterDelete` in order for the single delete not to introduce
more than 20% deletes, and keep the `TieredMergePolic
benwtrent commented on issue #11830:
URL: https://github.com/apache/lucene/issues/11830#issuecomment-1270205020
I was able to replicate @jtibshirani results.
Using the `glove-100-angular` from ann-benchmarks, here is the graph data
size:
```
Baseline:
154M _0_Lucene94Hn
jtibshirani commented on issue #11830:
URL: https://github.com/apache/lucene/issues/11830#issuecomment-1270359029
@benwtrent thanks for looking into this! To clarify, `OffHeapHnswGraph` will
always be used for searches. Usually it will be paged in from disk before heavy
searching, but it's
benwtrent commented on issue #11830:
URL: https://github.com/apache/lucene/issues/11830#issuecomment-1270415680
> Could you include a comparison of the search latency and recall numbers
you see with this approach? Sometimes with our benchmarks it's easy to miss
small differences in performa
jtibshirani commented on issue #11838:
URL: https://github.com/apache/lucene/issues/11838#issuecomment-1270564173
This is a great question! To me it'd make sense to try to move costly steps
into weight creation or scoring. It feels a little "off" to do the bulk of a
query's work during rewr
10 matches
Mail list logo
