Details
-
Bug
-
Status: Resolved
-
Blocker
-
Resolution: Fixed
-
0.10.2.0
-
None
-
3 Kafka broker machines and 3 kafka streams machines.
Each machine is Linux 64 bit, CentOS 6.5 with 64GB memory, 8 vCPUs running in AWS
31GB java heap space allocated to each KafkaStreams instance and 4GB allocated to each Kafka broker.
Description
Having multiple kafka streams application instances causes one or more instances to get get into file lock contention and the instance(s) become unresponsive with uncaught exception.
The exception is below:
22:14:37.621 [StreamThread-7] WARN o.a.k.s.p.internals.StreamThread - Unexpected state transition from RUNNING to NOT_RUNNING
22:14:37.621 [StreamThread-13] WARN o.a.k.s.p.internals.StreamThread - Unexpected state transition from RUNNING to NOT_RUNNING
22:14:37.623 [StreamThread-18] WARN o.a.k.s.p.internals.StreamThread - Unexpected state transition from RUNNING to NOT_RUNNING
22:14:37.625 [StreamThread-7] ERROR n.a.a.k.t.KStreamTopologyBase - Uncaught Exception:org.apache.kafka.streams.errors.ProcessorStateException: task directory [/data/kafka-streams/rtp-kstreams-metrics/0_119] doesn't exist and couldn't be created
at org.apache.kafka.streams.processor.internals.StateDirectory.directoryForTask(StateDirectory.java:75)
at org.apache.kafka.streams.processor.internals.StateDirectory.lock(StateDirectory.java:102)
at org.apache.kafka.streams.processor.internals.StateDirectory.cleanRemovedTasks(StateDirectory.java:205)
at org.apache.kafka.streams.processor.internals.StreamThread.maybeClean(StreamThread.java:753)
at org.apache.kafka.streams.processor.internals.StreamThread.runLoop(StreamThread.java:664)
at org.apache.kafka.streams.processor.internals.StreamThread.run(StreamThread.java:368)
This happens within couple of minutes after the instances are up and there is NO data being sent to the broker yet and the streams app is started with auto.offset.reset set to "latest".
Please note that there are no permissions or capacity issues. This may have nothing to do with number of instances, but I could easily reproduce it when I've 3 stream instances running. This is similar to the (and may be the same) bug as KAFKA-3758
Here are some relevant configuration info:
3 kafka brokers have one topic with 128 partitions and 1 replication
3 kafka streams applications (running on 3 machines) have a single processor topology and this processor is not doing anything (the process() method just returns and the punctuate method just commits)
There is no data flowing yet, so the process() and puctuate() methods are not even called yet.
The 3 kafka stream instances have 43, 43 and 42 threads each respectively (totally making up to 128 threads, so one task per thread distributed across three streams instances on 3 machines).
Here are the configurations that I'd played around with:
session.timeout.ms=300000
heartbeat.interval.ms=60000
max.poll.records=100
num.standby.replicas=1
commit.interval.ms=10000
poll.ms=100
When punctuate is scheduled to be called every 1000ms or 3000ms, the problem happens every time. If punctuate is scheduled for 5000ms, I didn't see the problem in my test scenario (described above), but it happened in my real application. But this may have nothing to do with the issue, since punctuate is not even called as there are no messages streaming through yet.
