tez-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jaume M (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (TEZ-3974) Tez: Correctness regression of TEZ-955 in TEZ-2937
Date Mon, 06 Aug 2018 15:44:00 GMT

    [ https://issues.apache.org/jira/browse/TEZ-3974?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16570380#comment-16570380
] 

Jaume M commented on TEZ-3974:
------------------------------

Sure, to reproduce this the following step where taken:

* Create a DAG with two mappers and one reducer:
A      B
 \    /
   C

* Purposely throw a RuntimeException in B in the close method of the processor. This accomplishes
the empty events being sent but the task not finishing correctly. The RuntimeException is
only thrown once, the second attempt succeeds.
* Tweak tez.shuffle-vertex-manager.min-src-fraction, tez.shuffle-vertex-manager.max-src-fraction
so C is scheduled even though A, B may not have finished yet.
* We see A start, B start, A finish, B throw the exception, C start, C finish, B finish. So
C finishes before B which shouldn't happen.

After the proposed path C finishes after B.

 

> Tez: Correctness regression of TEZ-955 in TEZ-2937
> --------------------------------------------------
>
>                 Key: TEZ-3974
>                 URL: https://issues.apache.org/jira/browse/TEZ-3974
>             Project: Apache Tez
>          Issue Type: Bug
>            Reporter: Gopal V
>            Assignee: Jaume M
>            Priority: Critical
>         Attachments: TEZ-3974.1.patch
>
>
> TEZ-2937 might have introduced a race condition for Tez output events, along with TEZ-2237
> {code}
>       // Close the Outputs.
>       for (OutputSpec outputSpec : outputSpecs) {
>         String destVertexName = outputSpec.getDestinationVertexName();
>         initializedOutputs.remove(destVertexName);
>         List<Event> closeOutputEvents = ((LogicalOutputFrameworkInterface)outputsMap.get(destVertexName)).close();
>         sendTaskGeneratedEvents(closeOutputEvents,
>             EventProducerConsumerType.OUTPUT, taskSpec.getVertexName(),
>             destVertexName, taskSpec.getTaskAttemptID());
>       }
>       // Close the Processor.
>       processorClosed = true;
>       processor.close();
> {code}
> As part of TEZ-2237, the outputs send empty events when the output is closed without
being started (which happens in task init failures).
> These events are obsoleted when a task fails and this happens in the AM, but not before
the dispatcher looks at them.
> Depending on the timing, the empty events can escape obsoletion & be sent to a downstream
task.
> This gets marked as a SKIPPED event in the downstream task, which means that further
obsoletion events sent to the downstream task is ignored (because a zero byte fetch is not
repeated on node failure).
> So the downstream task can exit without actually waiting for the retry of the failed
task and cause silent dataloss in case where the retry succeeds in another attempt.
> So if processor.close() throws an exception, this introduce a race condition and if the
AM is too fast, we end up with correctness issues.
> This was originally reported in TEZ-955



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Mime
View raw message