Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Plain View
Kafka, mail # user - ASyncProducer loss more messages than SyncProducer when broker crashes.


+
王国栋 2013-02-04, 06:01
+
Jun Rao 2013-02-04, 16:44
+
王国栋 2013-02-05, 03:12
+
Jun Rao 2013-02-05, 06:20
+
王国栋 2013-02-05, 06:31
Copy link to this message
-
Re: ASyncProducer loss more messages than SyncProducer when broker crashes.
Jun Rao 2013-02-05, 09:03
It could be. This is because it takes ZK up to 6 secs to detect the failure
due to a hard kill (since ZK client is not closed properly). You can shrink
ZK session timeout to reduce the delay.

Thanks,

Jun

On Mon, Feb 4, 2013 at 10:30 PM, 王国栋 <[EMAIL PROTECTED]> wrote:

> We kill the broker on purpose, since we are doing some fail over tests. If
> the broker is hard killed, will the async-producer lose message for 6
> seconds?
>
> Thanks.
>
> Guodong
>
>
> On Tue, Feb 5, 2013 at 2:19 PM, Jun Rao <[EMAIL PROTECTED]> wrote:
>
> > Is the broker shut down cleanly? If so, the producer should get a ZK
> > notification immediately. If the broker is hard killed, the ZK
> notification
> > won't arrive at the producer until zk session timeout (default is 6
> secs).
> >
> > Thanks,
> >
> > Jun
> >
> > On Mon, Feb 4, 2013 at 7:11 PM, 王国栋 <[EMAIL PROTECTED]> wrote:
> >
> > > Hi Jun
> > >
> > > We are using zookeeper. And the async-producer loses  messages in 'X'
> > > seconds. Here, X is between 4 and 6. It seems that X is comparable with
> > > zookeeper session timeout.
> > >
> > > I am wondering the behavior of async-producer when socket is broken,
> but
> > > the broker is still visible for zookeeper client.
> > >
> > >
> > > On Tue, Feb 5, 2013 at 12:39 AM, Jun Rao <[EMAIL PROTECTED]> wrote:
> > >
> > > > Currently, there is no callback to report socket failure in the async
> > > > producer. However, async producer should be able to detect socket
> > failure
> > > > and reconnect to another broker. Are you using a vip or Zookeeper in
> > the
> > > > producer for load balancing?
> > > >
> > > > Thanks,
> > > >
> > > > Jun
> > > >
> > > > On Sun, Feb 3, 2013 at 10:00 PM, 王国栋 <[EMAIL PROTECTED]> wrote:
> > > >
> > > > > Hi,
> > > > >
> > > > > we are doing some fail over test on Kafka0.7.2. We kill one of the
> > > > brokers
> > > > > on purpose. But we find that the async-producer loss more messages.
> > We
> > > do
> > > > > catch exceptions for send api. But it seems that the async-producer
> > > > doesn't
> > > > > throw any exceptions when broker is done.
> > > > >
> > > > > We change the producer to sync-producer, and got the exceptions as
> > soon
> > > > as
> > > > > the broker is down.
> > > > >
> > > > > So I am wondering if I can get the socket status connected to the
> > > broker.
> > > > > Does some callbacks can do this?
> > > > >
> > > > > Thanks.
> > > > > --
> > > > > Guodong Wang
> > > > > 王国栋
> > > > >
> > > >
> > >
> > >
> > >
> > > --
> > > Guodong Wang
> > > 王国栋
> > >
> >
>
>
>
> --
> Guodong Wang
> 王国栋
>

 
+
王国栋 2013-02-05, 12:55