[
https://issues.apache.org/jira/browse/LUCENE-10676?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17576678#comment-17576678
]
Michael McCandless commented on LUCENE-10676:
---------------------------------------------
Is each field name exotically long as well?
Lucene used to do just this -- intern {{FieldInfo.name}} and then use `==` to
compare field names everywhere. But we decided long ago that this was
dangerous and not an important optimization. Still, that decision was maybe
pre Java 7 days when the intern'd pool was stored in {{PermGen}} instead of
"ordinary" heap and was more likely to cause {{OutOfMemoryError}}?
Maybe dig into those long ago issues / dev list thread to see the motivation to
stop interning?
> FieldInfo#name contributes significantly to heap usage at scale
> ---------------------------------------------------------------
>
> Key: LUCENE-10676
> URL: https://issues.apache.org/jira/browse/LUCENE-10676
> Project: Lucene - Core
> Issue Type: Bug
> Components: core/codecs
> Affects Versions: 9.3
> Environment: Seen in Lucene 9.3.0 running on Linux using JDK18 but
> seems independent of environment.
> Reporter: David Turner
> Priority: Minor
> Labels: heap, scalability
>
> We encountered an Elasticsearch user with high heap usage, a significant
> proportion of which was down to the contents of `FieldInfo#name`.
> This user was certainly pushing some scalability boundaries: this single
> process had thousands of active Lucene indices, many with 10k+ fields, and
> many indices had hundreds of segments due to an excess of flushes, so in
> total they had an enormous number of `FieldInfo` instances. Still, the bulk
> of the heap usage was just field names, and the total number of distinct
> field names was fairly small. That's pretty common, especially for time-based
> data like logs. Some kind of interning or deduplication of these strings
> would have reduced their heap usage by many GBs.
> Is there a way we could deduplicate these strings? Deduplicating them across
> segments within each index would already have helped, but ideally we'd like
> to deduplicate them across indices too.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]