Ron Dagostino created KAFKA-12480:
-------------------------------------
Summary: Reuse bootstrap servers in clients when last alive broker
in cluster metadata is unavailable
Key: KAFKA-12480
URL: https://issues.apache.org/jira/browse/KAFKA-12480
Project: Kafka
Issue Type: Improvement
Components: clients
Reporter: Ron Dagostino
https://issues.apache.org/jira/browse/KAFKA-12455 documented how a Java client
can temporarily lose connectivity to a 2-broker cluster that is undergoing a
roll because the client will repeatedly retry connecting to the last alive
broker that it knows about in the cluster metadata even when that broker is
unavailable. The client could potentially fallback to its bootstrap brokers in
this case and reconnect to the cluster quicker.
For example, assume a 2-broker cluster has broker IDs 1 and 2 and both appear
in the bootstrap servers for a consumer. Assume broker 1 rolls such that the
Java consumer receives a new METADATA response and only knows about broker 2
being alive, and then broker 2 rolls before the consumer gets a new METADATA
response indicating that broker 1 is also alive. At this point the Java
consumer will keep retrying broker 2, and it will not reconnect to the cluster
unless/until broker 2 becomes available -- or the client itself is restarted so
it can use its bootstrap servers again. Another possibility is to fallback to
the full bootstrap servers list when the last alive broker becomes unavailable.
I believe librdkafka-based client may perform this fallback, though I am not
certain. We should consider it for Java clients.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
