[ https://issues.apache.org/jira/browse/KAFKA-2792?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14999211#comment-14999211 ]
Guozhang Wang commented on KAFKA-2792: -------------------------------------- [~ewencp] [~onurkaraman] If we fire-and-forget in the close() call, it will likely lead to an EOF exception on the server side whenever some consumer closes themselves, causing some log pollutions. This is why I was deciding to not do fire-and-forget in close() but only in unsubscribe(). We need to think about if this issue can be resolved in the socket server if we stick to this plan. > KafkaConsumer.close() can block unnecessarily due to leave group waiting for > a reply > ------------------------------------------------------------------------------------ > > Key: KAFKA-2792 > URL: https://issues.apache.org/jira/browse/KAFKA-2792 > Project: Kafka > Issue Type: Bug > Components: consumer > Reporter: Ewen Cheslack-Postava > Assignee: Ewen Cheslack-Postava > Priority: Blocker > Fix For: 0.9.0.0 > > > The current implementation of close() waits for a response to LeaveGroup. > However, if we have an outstanding rebalance in the works, this can cause the > close() operation to have to wait for the entire rebalance process to > complete, which is annoying since the goal is to get rid of the consumer > object anyway. This is at best surprising and at worst can cause unexpected > bugs due to close() taking excessively long -- this was found due to > exceeding timeouts unexpectedly causing other operations in Kafka Connect to > timeout. > Waiting for a response isn't necessary since as soon as the data is in the > TCP buffer, it'll be delivered to the broker. The client doesn't benefit at > all from seeing the close group. So we can instead just always send the > request -- This message was sent by Atlassian JIRA (v6.3.4#6332)