[ https://issues.apache.org/jira/browse/LUCENE-5646?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13990300#comment-13990300 ]
Robert Muir commented on LUCENE-5646: ------------------------------------- I have not thought about what we shoudl do here 4.8.1-wise. I cannot prove that this optimization corrupts data yet, but i feel we should either remove it, or fix it and test the crap out of it. If there is objection to removing it, I will try to think about a test that really hammers it (possibly would be quite slow). I have the feeling something bad might happen... > stored fields bulk merging doesn't quite work right > --------------------------------------------------- > > Key: LUCENE-5646 > URL: https://issues.apache.org/jira/browse/LUCENE-5646 > Project: Lucene - Core > Issue Type: Bug > Reporter: Robert Muir > Fix For: 4.9, 5.0 > > Attachments: LUCENE-5646.patch > > > from doing some profiling of merging: > CompressingStoredFieldsWriter has 3 codepaths (as i see it): > 1. optimized bulk copy (no deletions in chunk). In this case compressed data > is copied over. > 2. semi-optimized copy: in this case its optimized for an existing > storedfieldswriter, and it decompresses and recompresses doc-at-a-time around > any deleted docs in the chunk. > 3. ordinary merging > In my dataset, i only see #2 happening, never #1. The logic for determining > if we can do #1 seems to be: > {code} > onChunkBoundary && chunkSmallEnough && chunkLargeEnough && noDeletions > {code} > I think the logic for "chunkLargeEnough" is out of sync with the > MAX_DOCS_PER_CHUNK limit? e.g. instead of: > {code} > startOffsets[it.chunkDocs - 1] + it.lengths[it.chunkDocs - 1] >= chunkSize // > chunk is large enough > {code} > it should be something like: > {code} > (it.chunkDocs >= MAX_DOCUMENTS_PER_CHUNK || startOffsets[it.chunkDocs - 1] + > it.lengths[it.chunkDocs - 1] >= chunkSize) // chunk is large enough > {code} > But this only works "at first" then falls out of sync in my tests. Once this > happens, it never reverts back to #1 algorithm and sticks with #2. So its > still not quite right. > Maybe [~jpountz] knows off the top of his head... -- This message was sent by Atlassian JIRA (v6.2#6252) --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org