You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@tez.apache.org by "Bikas Saha (JIRA)" <ji...@apache.org> on 2014/05/30 00:58:01 UTC

[jira] [Comment Edited] (TEZ-1151) Vertex should stay in initializing state until custom vertex manager sets the parallelism

    [ https://issues.apache.org/jira/browse/TEZ-1151?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14013062#comment-14013062 ] 

Bikas Saha edited comment on TEZ-1151 at 5/29/14 10:56 PM:
-----------------------------------------------------------

The first patch is smaller but it seemed inconsistent with the state machine and would be susceptible to errors down the road. So I changed the approach even though it was more work :). The patch would have been cleaner if the root input initializer always got invoked with -1 tasks but the distributor option prevents better unification of code paths because setParallelism() is not called at that time. The bulk of this patch is to unify the initialization -> init logic in one place and use it for all 3 case 1) root input 2) 1-1 split and 3) custom vertex only split.

bq. Need to handle the case where RootinputInitializers are pending
In the patch, V_PARALLELISM_INITIALIZED is sent only for case 3 - custom vertex manager only. Thus if there are root input initializers then this event is not sent. Ideally I would have liked to unify this but root input init can be with or without -1 tasks and that makes it hard.

Next patch will address the review comments.



was (Author: bikassaha):
The first patch is smaller but it seemed inconsistent with the state machine and would be susceptible to errors down the road. So I changed the approach even though it was more work :). The patch would have been cleaner if the root input initializer always got invoked with -1 tasks but the distributor option prevents better unification of code paths because setParallelism() is not called at that time. The bulk of this patch is to unify the initialization->init logic in one place and use it for all 3 case 1) root input 2) 1-1 split and 3) custom vertex only split.

bq. Need to handle the case where RootinputInitializers are pending
In the patch, V_PARALLELISM_INITIALIZED is sent only for case 3 - custom vertex manager only. Thus if there are root input initializers then this event is not sent. Ideally I would have liked to unify this but root input init can be with or without -1 tasks and that makes it hard.

Next patch will address the review comments.


> Vertex should stay in initializing state until custom vertex manager sets the parallelism
> -----------------------------------------------------------------------------------------
>
>                 Key: TEZ-1151
>                 URL: https://issues.apache.org/jira/browse/TEZ-1151
>             Project: Apache Tez
>          Issue Type: Bug
>            Reporter: Daniel Dai
>            Assignee: Bikas Saha
>         Attachments: TEZ-1151.1.patch, TEZ-1151.2.patch, TEZ-1151.3.patch
>
>
> In TEZ-1145, application is allowed to set -1 parallelism on a non 1-1 input vertex, and expect VertexManager to set the actual parallelism dynamically. However, when I does that, I hit the following exception:
> {code}
> org.apache.tez.dag.api.TezUncheckedException: vertex_1400869346120_0005_1_04 has -1 tasks but neither input initializers nor 1-1 uninited sources
>         at org.apache.tez.dag.app.dag.impl.VertexImpl$InitTransition.handleInitEvent(VertexImpl.java:2485)
>         at org.apache.tez.dag.app.dag.impl.VertexImpl$InitTransition.transition(VertexImpl.java:2431)
>         at org.apache.tez.dag.app.dag.impl.VertexImpl$InitTransition.transition(VertexImpl.java:2412)
>         at org.apache.hadoop.yarn.state.StateMachineFactory$MultipleInternalArc.doTransition(StateMachineFactory.java:385)
>         at org.apache.hadoop.yarn.state.StateMachineFactory.doTransition(StateMachineFactory.java:302)
>         at org.apache.hadoop.yarn.state.StateMachineFactory.access$300(StateMachineFactory.java:46)
>         at org.apache.hadoop.yarn.state.StateMachineFactory$InternalStateMachine.doTransition(StateMachineFactory.java:448)
>         at org.apache.tez.dag.app.dag.impl.VertexImpl.handle(VertexImpl.java:1267)
>         at org.apache.tez.dag.app.dag.impl.VertexImpl.handle(VertexImpl.java:158)
>         at org.apache.tez.dag.app.DAGAppMaster$VertexEventDispatcher.handle(DAGAppMaster.java:1716)
>         at org.apache.tez.dag.app.DAGAppMaster$VertexEventDispatcher.handle(DAGAppMaster.java:1702)
>         at org.apache.hadoop.yarn.event.AsyncDispatcher.dispatch(AsyncDispatcher.java:134)
>         at org.apache.hadoop.yarn.event.AsyncDispatcher$1.run(AsyncDispatcher.java:81)
>         at java.lang.Thread.run(Thread.java:695)
> {code}



--
This message was sent by Atlassian JIRA
(v6.2#6252)