Yifan, Did you look into the producer metrics and check the avg. batch size / request size? Did they change significantly compared with 0.8.2.1 clients (there should be similar metrics but with different metrics name)?
Guozhang On Mon, Sep 5, 2016 at 11:17 PM, Yifan Ying <[email protected]> wrote: > Hi Ismael, > > Thanks for replying. > > Yes. It's the comparison between 0.8.2.1 and 0.10.0.0 clients connecting to > 0.10.0.0 brokers(log.message.format.version=0.8.2.1). No compression is > used and that's all the producer config. > > From the Kafka code, I think the request-latency-avg metric seem not > include the blocking time, but only measures the time between creating > requests and getting response from brokers. Blocking on full buffer happens > before that. Please correct me if I am wrong. > > Yifan > > > On Mon, Sep 5, 2016 at 5:59 AM, Ismael Juma <[email protected]> wrote: > > > One thing that changed in 0.9.0.0 is the introduction of `max.block.ms`, > > which is 60 seconds by default. Since you have > > "block.on.buffer.full=false", > > you may want to tweak `max.block.ms`. > > > > Ismael > > > > On Mon, Sep 5, 2016 at 1:54 PM, Ismael Juma <[email protected]> wrote: > > > > > Hi Yifan, > > > > > > Just to make sure I understand, the comparison is between 0.8.2.1 and > > > 0.10.0.0 clients (both) connecting to a 0.10.0.0 cluster? We haven't > > > received reports of increased average producer latency and I'm not > aware > > of > > > a change that could cause such a dramatic difference. Is compression > > being > > > used and did you paste the full producer config? > > > > > > Ismael > > > > > > On Fri, Sep 2, 2016 at 6:31 PM, Yifan Ying <[email protected]> wrote: > > > > > >> The load before and after the upgrade are pretty similar. And all > these > > >> configs stayed same too. > > >> > > >> Also please correct me if I am wrong. I think the request latency = > > >> response received time - request created time. And request created > time > > is > > >> created when the records are ready to be sent after 'lingering'. > > >> > > >> Yifan > > >> > > >> On Thu, Sep 1, 2016 at 11:01 PM, Gerard Klijs < > [email protected]> > > >> wrote: > > >> > > >> > With a linger of 5 seconds, 2-3 seconds would make sense when the > load > > >> is > > >> > smaller, are are sure the measurements with 0.8.2.1 where with the > > same > > >> > load and/or linger worked correctly there? > > >> > > > >> > On Fri, Sep 2, 2016 at 1:12 AM Yifan Ying <[email protected]> > wrote: > > >> > > > >> > > We tried to upgrade the Kafka clients dependency from 0.8.2.1 to > > >> > 0.10.0.0. > > >> > > As I understand, the producer client doesn't have major changes in > > >> 0.10, > > >> > so > > >> > > we kept the same producer config in the upgrade: > > >> > > > > >> > > retries 3 > > >> > > retry.backoff.ms 5000 > > >> > > timeout.ms 10000 > > >> > > block.on.buffer.full false > > >> > > linger.ms 5000 > > >> > > metadata.fetch.timeout.ms 30000 > > >> > > > > >> > > However, after the upgrade, the avg producer request latency > > >> increased by > > >> > > 5×, from a few hundreds milliseconds to 2-3 seconds. Anyone has > seen > > >> this > > >> > > issue? Is there any change in 0.9/0.10 that would cause this? > > >> > > > > >> > > Thanks! > > >> > > > > >> > > -- > > >> > > Yifan > > >> > > > > >> > > > >> > > >> > > >> > > >> -- > > >> Yifan > > >> > > > > > > > > > > > > -- > Yifan > -- -- Guozhang
