Hi folks- I work on a Lucene-based search system and we recently added Java Flight Recorder to our benchmark tooling. When looking through results, we found DataInput#skipBytes() to be a top contributor to garbage creation. We're using Lucene84SkipReader and always skipping over Impacts in our use-case. At first glance, it appeared pretty obvious that creating new instances of the skipBuffer byte[] for each instance of DataInput was the culprit.
It looks like alternatives were discussed originally in LUCENE-5583 <https://issues.amazon.com/issues/LUCENE-4947>, one of which being a thread-local implementation of the skip buffer (since it can't be a static field without breaking delegating subclasses, like ChecksumIndexInput). At the time, a thread-local was advised against <https://issues.apache.org/jira/browse/LUCENE-5583?focusedCommentId=13964258&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-13964258> by Uwe due to GC expense, but in our benchmarks, bringing in a thread-local implementation reduced overall GC time by ~7%. I'd like to revisit this implementation decision and discuss ways in which we can reduce this unnecessary garbage creation. It seems like moving to a thread-local implementation is a win here, but I'd love to hear more thoughts or alternative suggestions from the group. I'm new to this community, so I'm not sure the best way to proceed. Should I open a Jira issue as a next step? Thanks in advance! Cheers, -Greg
