Massimiliano Tomassi created CASSANDRA-14877:
------------------------------------------------
Summary: StreamCoordinator "leaks" threads
Key: CASSANDRA-14877
URL: https://issues.apache.org/jira/browse/CASSANDRA-14877
Project: Cassandra
Issue Type: Bug
Components: Streaming and Messaging
Reporter: Massimiliano Tomassi
Fix For: 2.1.21, 2.2.14, 3.0.18, 3.11.4, 4.0
Since Cassandra 2.1, streaming sessions are started by running a
StreamSessionConnector task for each session in a dedicated executor (a static
field of StreamCoordinator).
That executor is initialized with
DebuggableThreadPoolExecutor.createWithFixedPoolSize, which means that once
created (up to the given limit of the number of logical cores), its threads are
kept alive for Integer.MAX_VALUE seconds.
This practically means that once a node needs to establish streaming sessions
to n other nodes, it will create Math.min(n, numLogicalCores)
StreamConnectionEstablisher threads that will stay parked forever after
initializing (not completing) the session.
It seems preferable to replace
DebuggableThreadPoolExecutor.createWithFixedPoolSize with
DebuggableThreadPoolExecutor.createWithMaximumPoolSize which allows providing a
saner keep-alive period (e.g. a minute).
That's also what createWithFixedPoolSize's Javadoc recommends: If (most)
threads are expected to be idle most of the time, prefer createWithMaxSize()
instead.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
