If only we had some sort of system test framework with a producer performance test that we could parameterize with the different acks settings to validate these performance differences...
wrt out of order: yes, with > 1 in flight requests with retries, messages can get out of order. Becket had a great presentation addressing that and a bunch of other issues with no data loss pipelines: http://www.slideshare.net/JiangjieQin/no-data-loss-pipeline-with-apache-kafka-49753844 Short version: as things are today, you have to *really* understand the producer settings, and some producer internals, to get the exact behavior you want. On Mon, Jul 27, 2015 at 9:44 AM, Gwen Shapira <gshap...@cloudera.com> wrote: > Yeah, using acks=0 should result in higher throughput since we are not > limited by the roundtrip time to the broker. > > Btw. regarding in-flight requests: With acks = 1 (or -1), can we send > a message batch to a partition before the brokers "acked" a previous > request? Doesn't it risk getting messages out of order? > > On Mon, Jul 27, 2015 at 9:41 AM, Guozhang Wang <wangg...@gmail.com> wrote: > > I think there is still a subtle difference between "async with acks = 0" > > and "async with callback", that when the #.max-inflight-requests has > > reached the subsequent requests cannot be sent until previous responses > are > > returned (which could happen, for example, when the broker is slow / > > network issue happens) in the second case but not in the first. > > > > Given this difference, I feel there are still scenarios, though probably > > rare, that users would like to use "acks = 0" even with new producer's > > callbacks. > > > > Guozhang > > > > On Mon, Jul 27, 2015 at 9:25 AM, Mayuresh Gharat < > gharatmayures...@gmail.com > >> wrote: > > > >> So basically this means that with acks = 0, their is no guarantee that > the > >> message has been received by Kafka broker. I am just wondering, why > would > >> anyone be using acks = 0, since anyone using kafka and doing > >> producer.send() would want that, their message got to kafka brokers. > Also > >> as Jay said, with new producer with async mode, clients will not have to > >> wait for the response since it will be handled in callbacks. So the use > of > >> acks = 0 sounds very rare to me and I am not able to think of an usecase > >> around it. > >> > >> Thanks, > >> > >> Mayuresh > >> > >> On Sun, Jul 26, 2015 at 2:40 PM, Gwen Shapira <gshap...@cloudera.com> > >> wrote: > >> > >> > Aha! Yes, I was missing the part with the dummy response. > >> > Thank you! > >> > > >> > Gwen > >> > > >> > > >> > On Sun, Jul 26, 2015 at 2:17 PM, Ewen Cheslack-Postava > >> > <e...@confluent.io> wrote: > >> > > It's different because it changes whether the client waits for the > >> > response > >> > > from the broker at all. Take a look at > >> > NetworkClient.handleCompletedSends, > >> > > which fills in dummy responses when a response is not expected (and > >> that > >> > > flag gets set via Sender.produceRequest using acks != 0 as a flag to > >> > > ClientRequest). This means that the producer will invoke the > callback & > >> > > resolve the future as soon as the request hits the TCP buffer on the > >> > > client. At that point, the behavior of the broker wrt generating a > >> > response > >> > > doesn't matter -- the client isn't waiting on that response anyway. > >> > > > >> > > This definitely is faster since you aren't waiting for the round > trip, > >> > but > >> > > it seems like it is of questionable value with the new producer as > Jay > >> > > explained. It is slightly better than just assuming records have > been > >> > sent > >> > > as soon as you call Producer.send() in this shouldn't trigger a > >> callback > >> > > until the records have made it through the internal KafkaProducer > >> > > buffering. But since it still has to make it through the TCP > buffers it > >> > > doesn't really guarantee anything that useful. > >> > > > >> > > -Ewen > >> > > > >> > > > >> > > On Sun, Jul 26, 2015 at 1:40 PM, Gwen Shapira < > gshap...@cloudera.com> > >> > wrote: > >> > > > >> > >> What bugs me is that even with acks = 0, the broker will append to > >> > >> local log before responding (unless I'm misreading the code), so I > >> > >> don't see why a client with acks = 0 will be any faster. Unless the > >> > >> client chooses to not wait for response, which is orthogonal to > acks > >> > >> parameter. > >> > >> > >> > >> On Mon, Jul 20, 2015 at 8:52 AM, Jay Kreps <j...@confluent.io> > wrote: > >> > >> > acks=0 is a one-way send, the client doesn't need to wait on the > >> > >> response. > >> > >> > Whether this is useful sort of depends on the client > implementation. > >> > The > >> > >> > new java producer does all sends async so "waiting" on a response > >> > isn't > >> > >> > really a thing. For a client that lacks this, though, as some of > >> them > >> > do, > >> > >> > acks=0 will be a lot faster. > >> > >> > > >> > >> > It also makes some sense in terms of what is completed when the > >> > request > >> > >> is > >> > >> > considered satisfied > >> > >> > acks = 0 - message is written to the network (buffer) > >> > >> > acks = 1 - message is written to the leader log > >> > >> > acks = -1 - message is committed > >> > >> > > >> > >> > -Jay > >> > >> > > >> > >> > On Sat, Jul 18, 2015 at 10:50 PM, Gwen Shapira < > >> gshap...@cloudera.com > >> > > > >> > >> > wrote: > >> > >> > > >> > >> >> Hi, > >> > >> >> > >> > >> >> I was looking into the different between acks = 0 and acks = 1 > in > >> the > >> > >> >> new producer, and was a bit surprised at what I found. > >> > >> >> > >> > >> >> Basically, if I understand correctly, the only difference is > that > >> > with > >> > >> >> acks = 0, if the leader fails to append locally, it closes the > >> > network > >> > >> >> connection silently and with acks = 1, it sends an actual error > >> > >> >> message. > >> > >> >> > >> > >> >> Which seems to mean that with acks = 0, any failed produce will > >> lead > >> > >> >> to metadata refresh and a retry (because network error), while > >> acks = > >> > >> >> 1 will lead to either retries or abort, depending on the error. > >> > >> >> > >> > >> >> Not only this doesn't match the documentation, it doesn't even > make > >> > >> >> much sense... > >> > >> >> "acks = 0" was supposed to somehow makes things "less safe but > >> > >> >> faster", and it doesn't seem to be doing that any more. I'm not > >> even > >> > >> >> sure there's any case where the "acks = 0" behavior is > desirable. > >> > >> >> > >> > >> >> Is it my misunderstanding, or did we somehow screw up the logic > >> here? > >> > >> >> > >> > >> >> Gwen > >> > >> >> > >> > >> > >> > > > >> > > > >> > > > >> > > -- > >> > > Thanks, > >> > > Ewen > >> > > >> > >> > >> > >> -- > >> -Regards, > >> Mayuresh R. Gharat > >> (862) 250-7125 > >> > > > > > > > > -- > > -- Guozhang > -- Thanks, Ewen
