[
https://issues.apache.org/jira/browse/NIFI-16290?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18114905#comment-18114905
]
ASF subversion and git services commented on NIFI-16290:
--------------------------------------------------------
Commit d7c577293054be10c2dcaeb47ac15b4ee6a85aef in nifi's branch
refs/heads/main from Pierre Villard
[ https://gitbox.apache.org/repos/asf?p=nifi.git;h=d7c57729305 ]
NIFI-16290 - Node offload should give processors a bounded grace period to stop
cleanly before forced termination
This closes #11621
> Node offload should give processors a bounded grace period to stop cleanly
> before forced termination
> ----------------------------------------------------------------------------------------------------
>
> Key: NIFI-16290
> URL: https://issues.apache.org/jira/browse/NIFI-16290
> Project: Apache NiFi
> Issue Type: Improvement
> Components: Core Framework
> Reporter: Pierre Villard
> Assignee: Pierre Villard
> Priority: Major
> Time Spent: 40m
> Remaining Estimate: 0h
>
> Currently, StandardFlowService.offload() calls rootGroup.stopProcessing() and
> immediately discards the returned future, then force-terminates any processor
> whose logical scheduled state is STOPPED. Because logical STOPPED also covers
> the physical STOPPING state, this termination can fire before a processor's
> asynchronous stop lifecycle has actually finished — overtaking
> @OnUnscheduled, aborting the wait for active threads to drain, and bypassing
> @OnStopped entirely. As a result, processors that close external resources
> (connections, files, sessions) during shutdown may never get the chance to do
> so cleanly when a node is offloaded, even though nothing was actually wedged
> or slow.
> This change introduces a bounded grace period: offload now waits on the
> aggregate stopProcessing() future for up to
> nifi.flowcontroller.graceful.shutdown.period (reusing the existing
> graceful-shutdown setting) before falling through to the existing
> forced-termination sweep, which still always runs unconditionally afterward.
> This preserves the current guarantee that a hung or wedged processor —
> including one that has stashed an uncommitted ProcessSession, per NIFI-15885
> — cannot block node offload indefinitely, while giving well-behaved
> processors a real opportunity to complete @OnUnscheduled/@OnStopped normally
> first. The outcome of the graceful wait (completed, timed out, interrupted,
> or exceptional/cancelled) is logged for operational visibility.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
