Yes, this was observed to be a problem with
CompressedStoredFieldsReader, and there may indeed be other cases. The
new storedFields API should help as of Lucene 9.5, but I think we'd
still do well to also pursue a faster QueuedThreadPool shrinking
configuration.
The current Solr default jetty
Has anyone experimented with reducing the number of request/Jetty threads
in jetty.xml? The default is 10K. I'm concerned about the use of
ThreadLocals caching stuff, like maybe Lucene analysis chains (Analyzer
ReuseStrategy) or other things. I think Michael Gibney reported to the
Jetty project