[
https://issues.apache.org/jira/browse/LUCENE-7868?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Michael McCandless updated LUCENE-7868:
---------------------------------------
Attachment: LUCENE-7868.patch
Another iteration; tests seem to be passing consistently now, but I still have
27 nocommits.
I re-tested the speedup with this change, re-indexing nearly a billion
documents into an index that already had all of those documents indexed once,
and the overall speedup is ~53% faster with the patch (26.8 K docs/sec vs 17.5
K docs/sec).
I still need to test doc-values updates performance.
Besides the performance gains from concurrent deletes/updates, the patch has
two user-visible changes:
* Reader pooling is now enabled by default in IndexWriter; previously it was
only turned on the first time you pulled an NRT reader. You can still disable
this, but it will hurt indexing perf since all deletes/updates will be eagerly
written through to the filesystem.
* Removed IndexWriterConfig.set/getMaxBufferedDeleteTerms; this setting is no
longer possible because Lucene eagerly resolves the deletes and updates.
> Use multiple threads to apply deletes and DV updates
> ----------------------------------------------------
>
> Key: LUCENE-7868
> URL: https://issues.apache.org/jira/browse/LUCENE-7868
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Michael McCandless
> Assignee: Michael McCandless
> Fix For: master (7.0)
>
> Attachments: cpu-after.png, cpu-before.png, LUCENE-7868.patch,
> LUCENE-7868.patch
>
>
> Today, when users delete documents or apply doc values updates, IndexWriter
> buffers them up into frozen packets and then eventually uses a single thread
> (BufferedUpdatesStream.applyDeletesAndUpdates) to resolve delete/update terms
> to docids. This thread also holds IW's monitor lock, so it also blocks
> refresh, merges starting/finishing, commits, etc.
> We have heavily optimized this part of Lucene over time, e.g. LUCENE-6161,
> LUCENE-2897, LUCENE-2680, LUCENE-3342, but still, it's a single thread so it
> can't use multiple CPU cores commonly available now.
> This doesn't affect append-only usage, but for update-heavy users (me!) this
> can be a big bottleneck, and causes long stop-the-world hangs during indexing.
> I have an initial exploratory patch to make these lookups concurrent, without
> holding IW's lock, so that when a new packet of deletes is pushed, which
> happens when we flush a new segment, we immediately use that same indexing
> thread to and resolve the deletions.
> This is analogous to when we made segment flushing concurrent (LUCENE-3023),
> just for deletes and DV updates as well.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
