flink-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "ASF GitHub Bot (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (FLINK-3633) Job submission silently fails when using user code types
Date Mon, 21 Mar 2016 16:21:25 GMT

    [ https://issues.apache.org/jira/browse/FLINK-3633?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15204529#comment-15204529
] 

ASF GitHub Bot commented on FLINK-3633:
---------------------------------------

Github user tillrohrmann commented on the pull request:

    https://github.com/apache/flink/pull/1818#issuecomment-199363606
  
    Due to the tight coupling of the ExecutionConfig and multiple Flink components
    (e.g. PojoSerializer) the automatic serialization and manual deserialization of
    user code objects via the UserCodeValue class caused problems. In order to minimize
    the impact of the changes, I changed the serialization strategy to an explicit one.
    One has to call `ExecutionConfig.serializeUserCode` to store the user code objects in
a SerializedValue
    object and nulling the corresponding member fields. If that is not done, then it is
    assumed that the object is deserialized using a user code class loader. On the receiving
side one has to call `ExecutionConfig.deserializeConfig` providing a class loader which knows
the user code classes.


> Job submission silently fails when using user code types
> --------------------------------------------------------
>
>                 Key: FLINK-3633
>                 URL: https://issues.apache.org/jira/browse/FLINK-3633
>             Project: Flink
>          Issue Type: Bug
>    Affects Versions: 1.1.0
>            Reporter: Till Rohrmann
>            Assignee: Till Rohrmann
>            Priority: Blocker
>             Fix For: 1.1.0
>
>
> With the changes introduced by FLINK-3327, it is no longer possible to run remote Flink
jobs which work on user code types. The reason is that now the {{ExecutionConfig}} is directly
stored in the {{JobGraph}} which is sent as an Akka message to the {{JobManager}}. Per default,
user code types are automatically detected and registered in the {{ExecutionConfig}}. When
deserializing a {{JobGraph}} whose {{ExecutionConfig}} contains user code classes the user
code class loader is consequently required. However, Akka does not have access to it and uses
the system class loader. This causes that Akka silently discards the {{SubmitJob}} message
which cannot be deserialized because of a {{ClassNotFoundException}}.
> I propose to not sent the {{ExecutionConfig}} explicitly with the {{JobGraph}} and, thus,
to partially revert the changes to before FLINK-3327. Before, the {{ExectuionConfig}} was
serialized into the job configuration and deserialized on the {{TaskManager}} using the proper
user code class loader.
> In order to reproduce the problem you can submit the following job to a remote cluster.
> {code}
> public class Job {
> 	public static class CustomType {
> 		private final int value;
> 		public CustomType(int value) {
> 			this.value = value;
> 		}
> 		@Override
> 		public String toString() {
> 			return "CustomType(" + value + ")";
> 		}
> 	}
> 	public static void main(String[] args) throws Exception {
> 		ExecutionEnvironment env = ExecutionEnvironment.createRemoteEnvironment(Address, Port,
PathToJar);
> 		env.getConfig().disableAutoTypeRegistration();
> 		DataSet<Integer> input = env.fromElements(1,2,3,4,5);
> 		DataSet<CustomType> customTypes = input.map(new MapFunction<Integer, CustomType>()
{
> 			@Override
> 			public CustomType map(Integer integer) throws Exception {
> 				return new CustomType(integer);
> 			}
> 		});
> 		customTypes.print();
> 	}
> }
> {code}



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message