Hi Jason, Thanks for the KIP. I think this is a long-overdue improvement.
It would be good to clarify that the "timeout" values for the options classes will now refer to the API timeouts, not to the request timeouts. I think this is implied by the KIP, but it would be good to spell it out. One interesting question is if we should let users configure the request timeout on a per-request basis as well. I guess if we don't have a clear use-case, we could leave this as a possible future improvement. I'm a bit concerned that the proposed changes will lead to faulty clients spamming the servers. In particular, the default retry.backoff.ms is only 100 ms and you are now increasing the number of retries to be effectively unlimited. I remember we had a discussion offline about this and I commented that we already had exponential backoff, but I was misremembering. We actually only have it for reconnecting. To address this, I think we should add exponential backoff here, like we did with reconnect backoff. We could have a similar retry.backoff.max.ms to set the upper limit for backoff. best, Colin On Wed, Oct 9, 2019, at 12:06, Jason Gustafson wrote: > Hi All, > > I wrote a short KIP to address a longstanding issue with timeout > behavior > in the AdminClient: > https://cwiki.apache.org/confluence/display/KAFKA/KIP-533%3A+Add+default+api+timeout+to+AdminClient > . > > Take a look and let me know what you think. > > Thanks, > Jason >
