[
https://issues.apache.org/jira/browse/CAMEL-22410?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Claus Ibsen reassigned CAMEL-22410:
-----------------------------------
Assignee: Claus Ibsen
> SchedulingPollConsumer is not thread safe during graceful shutdown.
> -------------------------------------------------------------------
>
> Key: CAMEL-22410
> URL: https://issues.apache.org/jira/browse/CAMEL-22410
> Project: Camel
> Issue Type: Bug
> Affects Versions: 4.12.0
> Reporter: Ruben Lapauw
> Assignee: Claus Ibsen
> Priority: Minor
> Labels: sqs
> Fix For: 4.15.0
>
>
> SchedulingPollConsumer has a race condition that causes the SQS library to
> throw an SdkInterruptedException during shutdown. This results in timed-out
> messages delivered to the dead-letter-queue.
> The bug is rare when concurrentConsumers=1 but can be reliably triggered when
> concurrentConsumers>=2.
> h6. Replication:
> * Connect to
>
> {code:java}
> - route:
> id: "test-route"
> shutdownRunningTask: "CompleteAllTasks"
> from:
> uri:
> aws2-sqs://sqs-queue?concurrentConsumers=2&waitTimeSeconds=20&maxMessagesPerPoll=1
> steps:
> - log:
> message: Endpoint result is ${body} {code}
> * Trigger a shutdown.
> * Send three messages staggered two seconds apart.
> * Observe SdkInterruptedException for the second message
> h6. Cause:
> I am viewing the source code commit 7b766867.
> * For concurrentConsumers = 2 or more
> When concurrentConsumers = 2 then two threads poll at the same time. Thus
> when each starts a poll, both set the shared boolean variable 'polling' to
> true [line 203 of ScheduledPollConsumer]. The first to finish polling, e.g.
> by receiving sufficient messages, resets 'polling' to false [line 236]. This
> allows the DefaultShutdownStrategy to progress to interrupting the threads
> past the wait-loop checking for pendingInflightExchanges [lines 674, 782].
> The SQS library checks after receiving messages and before processing them
> whether the Thread was interrupted. (I consider this reasonable behaviour.)
> The end result is a race condition where messages time out after being
> received because they are not fully processed.
> * For concurrentConsumers = 1
> In ScheduledPollConsumer line 194 to line 201 there is a gap during which a
> thread can sleep while the DefaultShutdownStrategy can progress from
> deferring the Shutdown to waiting for inflight messages to interrupting the
> threads of the deferredConsumers.
> The same race condition is as before is possible though with reduced scope.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
