-Re: Transactional writing
Jun Rao 2013-03-28, 15:23
With a single writer, the producer can achieve exact once write. If a send
request fails, the producer first checks the end of the log to see if the
previous write succeeded or not. The producer will only resend if the
previous write fails.
To do this, the producer needs the offset of appended messages. In 0.8,
such offsets are not returned in our high level producer API yet. We plan
to extend our producer API post 0.8 to expose this information.
On Wed, Mar 27, 2013 at 2:41 PM, Jonathan Hodges <[EMAIL PROTECTED]> wrote:
> I know this is a really old thread, but it looked like the only pertinent
> one that came up when searching for ‘exactly once’ in the archives. I just
> want to confirm my understanding of the 0.8 version in that it still
> doesn’t completely support exactly once semantics. With the producer
> configured in sync mode and quorum commits there are still some edge case
> failure modes where the producer won’t receive the ack and resends the
> message(s). I think I read that the consumers don’t see uncommitted
> messages in the log, but I don’t think that addresses this producer case.
> Please correct me if I am missing something here.
> Don’t get me wrong we are very thankful for the 0.8 features. It offers by
> far the best message delivery guarantees out of the products we evaluated
> like Rabbit and ActiveMQ. We attempt to make are downstream consumer
> processes idempotent to mitigate this edge case, but it isn’t always
> feasible. Also the suggestion by Milind in this thread of using Storm for
> exactly once guarantees makes a lot of sense. Trident State seems to
> address this very issue (
> https://github.com/nathanmarz/storm/wiki/Trident-state) so we could just
> have it mediate our topics that required exactly once.
> On Sat, Nov 3, 2012 at 1:53 PM, Milind Parikh <[EMAIL PROTECTED]
> > Why wouldn't the storm approach provide semantics of exactly once
> > delivery? https://github.com/nathanmarz/storm
> > Nathan actually credits the Kafka_devs for the basic idea of transaction
> > persisting in one of his talks.
> > Regards
> > Milind
> > On Nov 3, 2012 11:51 AM, "Rohit Prasad" <[EMAIL PROTECTED]>
> > > I agree that this approach only prevents duplicate messages to
> > > from the Producer side. There needs to be a similar approach on the
> > > consumer side too. Using Zk can be one solution, or other non-ZK
> > > approaches.
> > >
> > > Even if Consumer reads none or all messages of a transaction. But that
> > does
> > > not solve the transaction problem yet. Because the business/application
> > > logic inside the Consumer thread may execute partially and fail. So it
> > > becomes tricky to decide the point when you want to say that you have
> > > "consumed" the message and increase consumption offset. If your
> > > thread is saving some value into DB/HDFS/etc, ideally you want this
> > > operation and consumption offset to be incremented atomically. Thats
> > it
> > > boils down to Application logic implementing transactions and dealing
> > with
> > > duplicates.
> > > Maybe a journalling or redo log approach on Consumer side can help
> > > such a system.
> > >
> > > It will be nice if eventually kafka can be a transport which provides
> > > "exactly once" semantics for message delivery. Then consumer threads
> > be
> > > sure that they receive messages once, and can build appln logic on top
> > > that.
> > >
> > > I have a use case similar to what Jay mentioned in a previous mail. I
> > want
> > > to do aggregation but want the aggregated data to be correct, possible
> > > avoiding duplicates incase of failures/crashes.
> > >
> > >
> > >
> > > On Fri, Nov 2, 2012 at 4:12 PM, Tom Brown <[EMAIL PROTECTED]>
> > >
> > > > That approach allows a producer to prevent duplicate messages to the
> > > > partition, but what about the consumer? In my case, I don't want the