Evan Pollan created KAFKA-5875:
----------------------------------
Summary: Consumer group repeatedly fails to join, even across JVM
restarts: BufferUnderFlowException reading the {{version}} field in the
consumer protocol header
Key: KAFKA-5875
URL: https://issues.apache.org/jira/browse/KAFKA-5875
Project: Kafka
Issue Type: Bug
Reporter: Evan Pollan
I've seen this maybe once a month in our large cluster Kubernetes-based Kafka
consumers & producers. Every once in a while a consumer in a Kubernetes "pod"
get this error trying to join a consumer group:
{code}
{"level":"INFO","@timestamp":"2017-09-12T13:45:42.173+0000","logger":"org.apache.kafka.common.utils.AppInfoParser","message":"Kafka
version : 0.11.0.0","exception":""}
{"level":"INFO","@timestamp":"2017-09-12T13:45:42.173+0000","logger":"org.apache.kafka.common.utils.AppInfoParser","message":"Kafka
commitId : cb8625948210849f","exception":""}
{"level":"INFO","@timestamp":"2017-09-12T13:45:42.178+0000","logger":"org.apache.kafka.clients.consumer.internals.ConsumerCoordinator","message":"Revoking
previously assigned partitions [] for group
conv-fetch-jobs-runner-for-internal","exception":""}
{"level":"INFO","@timestamp":"2017-09-12T13:45:42.178+0000","logger":"org.apache.kafka.clients.consumer.internals.AbstractCoordinator","message":"(Re-)joining
group conv-fetch-jobs-runner-for-internal","exception":""}
{"level":"INFO","@timestamp":"2017-09-12T13:45:43.588+0000","logger":"org.apache.kafka.clients.consumer.internals.AbstractCoordinator","message":"Successfully
joined group conv-fetch-jobs-runner-for-internal with generation
17297","exception":""}
{"errorType":"Error reading field 'version':
java.nio.BufferUnderflowException","level":"ERROR","message":"Died!","operation":"Died!","stacktrace":"org.apache.kafka.common.protocol.types.SchemaException:
Error reading field 'version': java.nio.BufferUnderflowException\n\tat
org.apache.kafka.common.protocol.types.Schema.read(Schema.java:75)\n\tat
org.apache.kafka.clients.consumer.internals.ConsumerProtocol.deserializeAssignment(ConsumerProtocol.java:105)\n\tat
org.apache.kafka.clients.consumer.internals.ConsumerCoordinator.onJoinComplete(ConsumerCoordinator.java:220)\n\tat
org.apache.kafka.clients.consumer.internals.AbstractCoordinator.joinGroupIfNeeded(AbstractCoordinator.java:363)\n\tat
org.apache.kafka.clients.consumer.internals.AbstractCoordinator.ensureActiveGroup(AbstractCoordinator.java:310)\n\tat
org.apache.kafka.clients.consumer.internals.ConsumerCoordinator.poll(ConsumerCoordinator.java:297)\n\tat
org.apache.kafka.clients.consumer.KafkaConsumer.pollOnce(KafkaConsumer.java:1078)\n\tat
org.apache.kafka.clients.consumer.KafkaConsumer.poll(KafkaConsumer.java:1043)\n\tat
com.spredfast.common.kafka.consumer.RunnableConsumer.pollOnce(RunnableConsumer.java:141)\n\tat
com.spredfast.common.kafka.consumer.RunnableConsumer.access$000(RunnableConsumer.java:28)\n\tat
com.spredfast.common.kafka.consumer.RunnableConsumer$Loop.iterate(RunnableConsumer.java:125)\n\tat
com.spredfast.common.kafka.consumer.RunnableConsumer.run(RunnableConsumer.java:78)\n\tat
java.lang.Thread.run(Thread.java:745)\n","trackingId":"dead-consumer","logger":"com.spredfast.common.kafka.consumer.RunnableConsumer","loggingVersion":"UNKNOWN","component":"MetricsFetch","pid":25,"host":"fetch-2420457278-sh4f5","@timestamp":"2017-09-12T13:45:43.613Z"}
{code}
Pardon the log format -- these get sucked into logstash, thus the JSON.
Here's the raw stacktrace:
{code}
org.apache.kafka.common.protocol.types.SchemaException: Error reading field
'version': java.nio.BufferUnderflowException
at org.apache.kafka.common.protocol.types.Schema.read(Schema.java:75)
at
org.apache.kafka.clients.consumer.internals.ConsumerProtocol.deserializeAssignment(ConsumerProtocol.java:105)
at
org.apache.kafka.clients.consumer.internals.ConsumerCoordinator.onJoinComplete(ConsumerCoordinator.java:220)
at
org.apache.kafka.clients.consumer.internals.AbstractCoordinator.joinGroupIfNeeded(AbstractCoordinator.java:363)
at
org.apache.kafka.clients.consumer.internals.AbstractCoordinator.ensureActiveGroup(AbstractCoordinator.java:310)
at
org.apache.kafka.clients.consumer.internals.ConsumerCoordinator.poll(ConsumerCoordinator.java:297)
at
org.apache.kafka.clients.consumer.KafkaConsumer.pollOnce(KafkaConsumer.java:1078)
at
org.apache.kafka.clients.consumer.KafkaConsumer.poll(KafkaConsumer.java:1043)
at
com.spredfast.common.kafka.consumer.RunnableConsumer.pollOnce(RunnableConsumer.java:141)
at
com.spredfast.common.kafka.consumer.RunnableConsumer.access$000(RunnableConsumer.java:28)
at
com.spredfast.common.kafka.consumer.RunnableConsumer$Loop.iterate(RunnableConsumer.java:125)
at
com.spredfast.common.kafka.consumer.RunnableConsumer.run(RunnableConsumer.java:78)
at java.lang.Thread.run(Thread.java:745)
{code}
What's fascinating about this is:
* We have a liveness probe (Kubernetes term for a healthcheck whose failure
will cause the container backing the "pod" to be killed and restarted) attached
to the existence of dead consumers. When this situation happens, it _never_
resolves itself. Today, I found a pod that had been restarted 1023 times due to
this error.
* The only way to make it go away is to _delete_ the Kubernetes pod. This
causes it to be replaced by a pod on another Kubernetes host ("minion") using
the same docker image and configuration. Invariably, this pod comes up and all
consumers join just fine
* We have _not_ tried restarting the brokers when this happens.
* There must be something about the pod, container, or Kubernetes host that is
consistent across pod crash loops that factors into the consumer group join
process -- MAC? hostname? Can't be anything that is recomputed on JVM
restart, though...
Seems like there's either:
# a bug in the client (i.e. in its assumption that it can deserialize a
protocol header on successful return of that join future). maybe there's a
flavor of broker response that doesn't include this header?
# a bug in the broker in that it's sending an empty or undersized response to a
group join command in some situtations.
It's worth noting that the severity of this issue is magnified by the fact that
it requires manual intervention. It wouldn't be so bad if our healthcheck
failure tripped a pod restart, and the new JVM's consumers would join OK. But,
the fact that even a JVM restart doesn't do it means most resiliency plays
won't work.
BTW, I see a similar schema read failure in
https://issues.apache.org/jira/browse/KAFKA-4349, although the client code is
completely different (admin {{ConsumerGroupCommand}})
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)