Roman Puchkovskiy created IGNITE-17787:
------------------------------------------
Summary: Fix node failures due to
java.nio.channels.CancelledKeyException
Key: IGNITE-17787
URL: https://issues.apache.org/jira/browse/IGNITE-17787
Project: Ignite
Issue Type: Bug
Components: networking
Reporter: Roman Puchkovskiy
Assignee: Roman Puchkovskiy
Fix For: 2.15
In GridNioServer, sessions may be moved between workers to balance the load.
When a session is moved from worker A, the corresponding SelectionKey is
cancelled. When a session is moved back, the selector is registered with the
corresponding channel.
Key cancellation is an async procedure. It gets really executed during
selection operation (select() or selectNow()). It might happen that a session
is moved from worker A (key cancellation is requested), then it gets moved back
to worker A, but the worker did not have a chance to invoke select() (due to
contention or some other reason). In such a case, we'll try to re-register a
selector with a channel for which a key is still registered, which will cause a
CancelledKeyException to be thrown (which may crash the node if a corresponding
failure handler is used).
We can fix this by invoking selectNow() just after requesting key cancellation.
This will make sure the cancellation request is actually handled before the
session lands to another worker.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)