Rajini Sivaram created KAFKA-10520:
--------------------------------------
Summary: InitProducerId may be blocked if least loaded node is not
ready to send
Key: KAFKA-10520
URL: https://issues.apache.org/jira/browse/KAFKA-10520
Project: Kafka
Issue Type: Bug
Components: producer
Reporter: Rajini Sivaram
Fix For: 2.7.0
>From the logs of a failing producer that shows InitProducerId timing out after
>request timeout, it looks like we don't poll while waiting for transactional
>producer to be initialized and FindCoordinator request cannot be sent. The
>producer configuration used one bootstrap server and
>{color:#172b4d}`{color}{color:#067d17}{color:#172b4d}max.in.flight.requests.per.connection=1`.
> The failing sequence:{color}{color}
# {color:#067d17}{color:#172b4d}Producer sends MetadataRequest to least loaded
node (bootstrap server){color}
{color}
# {color:#067d17}{color:#172b4d}Producer is ready to send InitProducerId,
needs to find transaction coordinator{color}{color}
# {color:#067d17}{color:#172b4d}Producer creates FindCoordinator request, but
the only node known is the bootstrap server. Producer cannot send to this node
since there is already the Metadata request in flight and max.inflight is
1.{color}{color}
# {color:#067d17}{color:#172b4d}Producer waits without polling, so Metadata
response is not processed. InitProducerId times out eventually.{color}{color}
{color:#067d17}{color:#172b4d}We need to update the condition used to determine
whether Sender should poll() to fix this issue.{color}{color}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
