[ 
https://issues.apache.org/jira/browse/FLINK-6018?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15927612#comment-15927612
 ] 

ASF GitHub Bot commented on FLINK-6018:
---------------------------------------

Github user sunjincheng121 commented on the issue:

    https://github.com/apache/flink/pull/3534
  
    Hi, @StephanEwen thanks for your attention to this JIRA. This issue only a 
code cleanup. In fact this is not a bug. When I work on 
[FLINK-5995](https://issues.apache.org/jira/browse/FLINK-5995), when using 
`DefaultOperatorStateBackend#getOperatorState` got an exception, Then I checked 
out the `AbstractKeyedStateBackend#getPartitionedState`, Then find the code 
duplicate check. Just like description of this JIRA. 
    About `AbstractKeyedStateBackend#getOrCreateKeyedState` I'm not sure, but I 
suggest do the same thing with  `getPartitionedState`. Because I had check 
`WindowOperator's windowStateDescriptor` also use 
`input.getType().createSerializer(getExecutionEnvironment().getConfig())` init 
the `typeSerializer`. 
    What do you think? 


> Properly initialise StateDescriptor in 
> AbstractStateBackend.getPartitionedState()
> ---------------------------------------------------------------------------------
>
>                 Key: FLINK-6018
>                 URL: https://issues.apache.org/jira/browse/FLINK-6018
>             Project: Flink
>          Issue Type: Improvement
>          Components: DataStream API, State Backends, Checkpointing
>            Reporter: sunjincheng
>            Assignee: sunjincheng
>             Fix For: 1.3.0
>
>
> The code snippet currently in the `AbstractKeyedStateBackend # 
> getPartitionedState` method, as follows:
> {code}
> line 352: // TODO: This is wrong, it should throw an exception that the 
> initialization has not properly happened
> line 353: if (!stateDescriptor.isSerializerInitialized()) {
> line 354:        stateDescriptor.initializeSerializerUnlessSet(new 
> ExecutionConfig());
> line 354 }
> {code}
> Method `isSerializerInitialized`:
> {code}
> public boolean isSerializerInitialized() {
>               return serializer != null;
>       }
> {code}
> Method `initializeSerializerUnlessSet`:
> {code}
> public void initializeSerializerUnlessSet(ExecutionConfig executionConfig) {
>               if (serializer == null) { 
>                       if (typeInfo != null) {
>                               serializer = 
> typeInfo.createSerializer(executionConfig);
>                       } else {
>                               throw new IllegalStateException(
>                                               "Cannot initialize serializer 
> after TypeInformation was dropped during serialization");
>                       }
>               }
>       }
> {code}
> that is, in the `initializeSerializerUnlessSet` method, The `serializer` has 
> been checked by `serializer == null`.So I hope this code has a little 
> improvement to the following:
> approach 1: 
> According to the `TODO` information  we throw an exception
> {code}
> if (!stateDescriptor.isSerializerInitialized()) {
>                       throw new IllegalStateException("The serializer of the 
> descriptor has not been initialized!"); 
> }
> {code}
> approach 2:
> Try to initialize and remove `if (!stateDescriptor.isSerializerInitialized()) 
> {` logic.
> {code}
> stateDescriptor.initializeSerializerUnlessSet(new ExecutionConfig());
> {code}
> Meanwhile, If we use the approach 2, I suggest that 
> `AbstractKeyedStateBackend` add a `private final ExecutionConfig 
> executionConfig` property. then we can change the code like this:
> {code}
> stateDescriptor.initializeSerializerUnlessSet(executionConfig);
> {code}
> Are the above suggestions reasonable for you? 
> Welcome anybody's feedback and corrections.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

Reply via email to