Caleb Rackliffe created CASSANDRA-20668:
-------------------------------------------
Summary: SAI range queries can spend too much time in
TrieMemoryIndex$Collector#updateLastQueueSize() doing unproductive work
Key: CASSANDRA-20668
URL: https://issues.apache.org/jira/browse/CASSANDRA-20668
Project: Apache Cassandra
Issue Type: Improvement
Components: Feature/SAI
Reporter: Caleb Rackliffe
Assignee: Caleb Rackliffe
Attachments: e20ea330-3519-11f0-912e-31b7d282b812-flamegraph.html
{{TrieMemoryIndex}} performs range queries by traversing the in-memory trie,
finding the terms in the range, and adding the primary keys for each term to a
priority queue (because they need to be in order). At one point, to make sure
we didn't have an issue with the priority queue's backing array being
constantly resized for large ranges, we started tracking the last queue size to
reuse it when creating a new {{Collector}} for the query. In some recent
performance testing, we've seen up to 10% of samples in a CPU flame graph
(attached) in {{updateLastQueueSize()}}, which needs to be addressed.
The naïve question here is, of course, why the priority queue itself isn't
thread local. There are also some usages of lambdas and other
micro-optimizations that might be possible...
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
