Cristian created FLINK-11408:
--------------------------------

             Summary: ContinuousProcessingTimeTrigger: NPE on clear() and state 
is lost on merge
                 Key: FLINK-11408
                 URL: https://issues.apache.org/jira/browse/FLINK-11408
             Project: Flink
          Issue Type: Bug
         Environment: Put both bugs in 
[https://github.com/casidiablo/flink-continuous-processing-trigger-bugs]

This is running Flink 1.7.1 locally.
            Reporter: Cristian


I was testing session windows using processing time and found a couple of 
problems with the 

ContinuousProcessingTimeTrigger. The first one is an NPE in the clear method:

[https://github.com/casidiablo/flink-continuous-processing-trigger-bugs/blob/master/src/main/java/flink/bug/Bug1.java]

The second one, which is most likely related and the root cause of the first 
one, is that the way the state is merged for windows that are merged somehow 
makes it so that the trigger gets confused and it stops triggering:

 

[https://github.com/casidiablo/flink-continuous-processing-trigger-bugs/blob/master/src/main/java/flink/bug/Bug2.java]

 

I managed to solve both of these by using a modified version of 

ContinuousProcessingTimeTrigger which does NOT call 
`ctx.mergePartitionedState(stateDesc);` in the onMerge method.

This is what I understand it happens at the trigger level:
 * The first element in the stream sets an initial fire time (logic is in

ContinuousProcessingTimeTrigger#onElement()), if there is no trigger set.
 * From then on, ContinuousProcessingTimeTrigger#onProcessingTime() is 
responsible for scheduling the next trigger. 
 * When the windows are merged (in the case of session windows), somehow the 
clear and merge methods are called using the wrong window namespace (I think 
this is the root cause of the bug, but I'm not too familiar with that code).
 * Because the state is not cleared properly in the new window namespace, the 
previously scheduled trigger gets executed against the window that was cleared.
 * Moreover, the new window has the state of the previous window, which means 
that:
 ## onElement will NOT schedule a fire trigger
 ## onProcessingTime will never be called at all



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

Reply via email to