[APEXMALHAR-2493] KafkaSinglePortExactlyOnceOutputOperator going to the blocked state during recovery

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 3.8.0
• Assignee: Chaitanya Chebolu
• Reporter: Chaitanya Chebolu
• Created: 18/May/17 09:43
• Updated: 11/Nov/17 11:25
• Resolved: 22/May/17 19:32

Description

Steps to reproduce the issue:
---------------------------------------

• Created the Kafka topic with single partition.
• Created the application with the following DAG:
  BatchSequenceGenerator -> KafkaSinglePortExactlyOnceOutputOperator

1. of partitions of KafkaSinglePortExactlyOnceOutputOperator = 2. Let's say KO1, KO2 are the two instances.

• Launched the app, after some time, manually killed the one of the instance of "KafkaSinglePortExactlyOnceOutputOperator" operator(KO2).
• During recovery, the instance comes up and after some time, it goes to the blocked state. App master killed this instance.

Observation:
----------------

• There is an infinite while loop in rebuildPartialWindow() method.
• While loop will break on the below 2 conditions:
  a) # of trails for "polled records from Kafka is empty" = 10
  b) Crossed boundary (consumerRecord.offset() >= currentOffset)

In this scenario, KO1 keeps on writing the data to Kafka. So, the first condition will not satisfy.
Operator is not checking the 2nd condition because of the below continue statement:
if (!doesKeyBelongsToThisInstance(operatorId, consumerRecord.key()))

{ continue; }

Solution: First check the cross boundary condition and then check the doesKeyBelongsToThisInstance(..).

Issue Links

• links to: GitHub Pull Request #622