Kuan Po Tseng created KAFKA-21286:
-------------------------------------
Summary: Share consumer may delay fatal startup errors while
waiting for the first heartbeat
Key: KAFKA-21286
URL: https://issues.apache.org/jira/browse/KAFKA-21286
Project: Kafka
Issue Type: Bug
Components: clients
Reporter: Kuan Po Tseng
Assignee: Kuan Po Tseng
KAFKA-19804 initializes the heartbeat interval to request.timeout.ms to prevent
busy polling while the first heartbeat is in flight. However, this can delay
fatal error reporting in the share consumer.
ShareConsumerImpl.pollForFetches() waits for up to the minimum of
maximumTimeToWait() and the remaining poll timeout. If the first heartbeat
returns a fatal error, such as GROUP_AUTHORIZATION_FAILED, the network thread
enqueues an error event, but BackgroundEventHandler does not wake the fetch
buffer. The application thread processes the error only after its current
fetch-buffer wait finishes.
With the default request.timeout.ms and a sufficiently long poll timeout, this
can delay reporting an already received error by approximately 30 seconds.
AsyncKafkaConsumer already caps the fetch-buffer wait at retryBackoffMs when no
partitions are assigned. The share consumer has no equivalent cap.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
