[
https://issues.apache.org/jira/browse/KAFKA-1252?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13903603#comment-13903603
]
Jay Kreps commented on KAFKA-1252:
----------------------------------
One question to consider is how we want to handle retry backoff. Currently this
code retries immediately. There are some exceptions like TimeoutException and
NotLeaderForPartition that should actually retry immediately and others like
LeaderNotAvailableException where you might want to wait a bit.
If you think about it I'm not really sure if the user actually cares about the
number of retries. Currently they care whether retries are possible because
retries implies possible duplicates/loss. If we fixed this then I suspect you
would always want to retry within the timeout you had specified. To avoid
banging the server it might be good to backoff, but this is ultimately not
really the concern of the client.
The current implementation is simple, though. We might want to just take that
as is.
The ideal implementation would be (1) make retries idempotent, (2) always
retry, (3) have an "exponential" backoff of something like 0, 1, 2, 4, 8, 16,
etc to avoid killing the server, (4) bound this by the produce time limit the
client specifies.
> Implement re-tries in new producer
> ----------------------------------
>
> Key: KAFKA-1252
> URL: https://issues.apache.org/jira/browse/KAFKA-1252
> Project: Kafka
> Issue Type: Sub-task
> Components: producer
> Reporter: Jay Kreps
> Assignee: Jay Kreps
> Attachments: KAFKA-1252.patch
>
>
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
