Matt Byrd created CASSANDRA-20480:
-------------------------------------
Summary: Avoid system keyspace truncation record access blocking
on interval tree builds
Key: CASSANDRA-20480
URL: https://issues.apache.org/jira/browse/CASSANDRA-20480
Project: Apache Cassandra
Issue Type: Improvement
Reporter: Matt Byrd
Assignee: Matt Byrd
A call to SystemKeyspace::removeEndpoint may block for extended periods of time
behind an interval tree build since it does a blocking flush, explained in a
bit more detail here:
https://issues.apache.org/jira/browse/CASSANDRA-19597
This in turn takes the synchronized monitor lock on SystemKeyspace.class and
blocks various other methods from executing, including in particular
SystemKeyspace::getTruncationRecord, which is called on various hot paths to
determine whether to drop/filter mutations, (Commitlog replay, Hints, Batchlog
and also committing Paxos transactions)
This can cause a continuous stream of timeouts while operations are on-going (
which are making calls to removeEndpoint).
The vast majority of the time the truncation record would not have changed,
since it's only updated when one truncates which is typically rare.
If we factor out the synchronization of the truncation record to a monitor on a
static field, we avoid this problem, whilst still preventing various
concurrency problems when operating on the SystemKeyspace field
truncationRecords.
--
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
