Uploaded image for project: 'Kafka'
  1. Kafka
  2. KAFKA-13887

Running multiple instance of same stateful KafkaStreams application on single host raise Exception



    • Improvement
    • Status: Open
    • Minor
    • Resolution: Unresolved
    • 2.6.0
    • None
    • streams
    • None


      KAFKA-10716 locks the state store directory on the running host, as it stores the processId in a kafka-streams-process-metadata file in this path. As a result to run multiple instances of the same application on a single host each instance must run with different state.dir config, otherwise the following exception will be raised for the second instance:
      Exception in thread "main" org.apache.kafka.streams.errors.StreamsException: Unable to initialize state, this can happen if multiple instances of Kafka Streams are running in the same state directory
      at org.apache.kafka.streams.processor.internals.StateDirectory.initializeProcessId(StateDirectory.java:191)
      at org.apache.kafka.streams.KafkaStreams.<init>(KafkaStreams.java:868)
      at org.apache.kafka.streams.KafkaStreams.<init>(KafkaStreams.java:851)
      at org.apache.kafka.streams.KafkaStreams.<init>(KafkaStreams.java:821)
      at org.apache.kafka.streams.KafkaStreams.<init>(KafkaStreams.java:733)
      The easiest solution multi-threading. Running single instance with multiple threads, but the multi-threading programming is not suitable for all scenarios. e.g., when the tasks are CPU intensive, or in large scale scenarios, or fully utilizing multi core CPUS.
      The second solution is multi-processing. This solution on a single host needs extra work and advisor, as each instance needs to be run with different state.dir. It is a good enhancement if kafkaStreams could handle this config for multi instance.
      The proposed solution is that the KafkaStreams use the /{state.dir}/{application.id}/{ordinal.number} path instead of /{state.dir}/{application.id} to store the meta file and states. The ordinal.number starts with 0 and is incremental.
      When an instance starts it checks the ordinal.number directories start by 0 and finds the first subdirectory that is not locked and use that for its state directory, this way all the tasks assigns correctly on rebalance and multiple instance can be run on single host.




            Unassigned Unassigned
            simsinak Sina Askarnejad
            0 Vote for this issue
            2 Start watching this issue