Kafka, mail # user - Number of consumers limited to number of partitions

Number of consumers limited to number of partitions

Vaibhav Puranik 2012-10-11, 15:20

Hi all,

We have a topic with three partitions - one on each broker (Kafka 0.7.1).
Using ConsumerConnector I noticed that I can consume messages from three
consumers at most. When I start fourth one, one of the older one stops.

Is this as designed? If I want more parallelism, do I need to increase
number of partitions?

Regards,
Vaibhav
GumGum

Re: Number of consumers limited to number of partitions

Neha Narkhede 2012-10-11, 15:27

Hi Vaibhav,

Yes, that is the behavior by design. The parallelism on the consumers
is limited by the number of partitions for the topics that they are
subscribed to.

Thanks,
Neha

On Thu, Oct 11, 2012 at 8:20 AM, Vaibhav Puranik <[EMAIL PROTECTED]> wrote:
> Hi all,
>
> We have a topic with three partitions - one on each broker (Kafka 0.7.1).
> Using ConsumerConnector I noticed that I can consume messages from three
> consumers at most. When I start fourth one, one of the older one stops.
>
> Is this as designed? If I want more parallelism, do I need to increase
> number of partitions?
>
> Regards,
> Vaibhav
> GumGum

Re: Number of consumers limited to number of partitions

Jay Kreps 2012-10-11, 15:41

Yes, the reason for this is that if two consumers consumed the same
partition internally we would have to keep track of per-message consumption
(which is expensive) and the ordering guarantee would be broken (since the
actual processing in the two consumer processes might happen out of order).
Since you can always have more partitions this should not be a limitation,
just a parameter to be aware of.

-Jay

On Thu, Oct 11, 2012 at 8:20 AM, Vaibhav Puranik <[EMAIL PROTECTED]> wrote:

> Hi all,
>
> We have a topic with three partitions - one on each broker (Kafka 0.7.1).
> Using ConsumerConnector I noticed that I can consume messages from three
> consumers at most. When I start fourth one, one of the older one stops.
>
> Is this as designed? If I want more parallelism, do I need to increase
> number of partitions?
>
> Regards,
> Vaibhav
> GumGum
>

Re: Number of consumers limited to number of partitions

Vaibhav Puranik 2012-10-11, 16:34

Thanks Neha and Jay.
It makes sense.

Regards,
Vaibhav

On Thu, Oct 11, 2012 at 8:41 AM, Jay Kreps <[EMAIL PROTECTED]> wrote:

> Yes, the reason for this is that if two consumers consumed the same
> partition internally we would have to keep track of per-message consumption
> (which is expensive) and the ordering guarantee would be broken (since the
> actual processing in the two consumer processes might happen out of order).
> Since you can always have more partitions this should not be a limitation,
> just a parameter to be aware of.
>
> -Jay
>
> On Thu, Oct 11, 2012 at 8:20 AM, Vaibhav Puranik <[EMAIL PROTECTED]>
> wrote:
>
> > Hi all,
> >
> > We have a topic with three partitions - one on each broker (Kafka 0.7.1).
> > Using ConsumerConnector I noticed that I can consume messages from three
> > consumers at most. When I start fourth one, one of the older one stops.
> >
> > Is this as designed? If I want more parallelism, do I need to increase
> > number of partitions?
> >
> > Regards,
> > Vaibhav
> > GumGum
> >
>