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

Switch to Threaded View
Kafka >> mail # user >> Re: Hardware Loadbalancer vs zookeeper for producers


Copy link to this message
-
Re: Hardware Loadbalancer vs zookeeper for producers
>> So my question is if we go with a hardware load balancer, do all the
broker nodes have to treated equally? Ie: All broker nodes will have the
same topics and number of partitions for each topic?

All the brokers behind the same hardware load balancer or virtual IP will
be treated equally.

>> Or is it possible to have brokers with different topics? If we go with
this configuration how can the hardware load balancer direct the messages
to the right broker for a given topic? We were thinking that we might be
able to do this by having a different VIP's per topic. How do you guys do
it?

It is tricky to selectively host topics on a subset of brokers in 0.7. You
need to create the right directory structure for every topic/partition and
bounce the broker. But I'm wondering why you would want to selectively host
topics on certain brokers. Is the purpose being able to host more topics on
the same cluster? In 0.7, an easier strategy is to logically categorize
classes of topics and host topics from the same class on a separate
cluster. But this strategy of course only works if there are fewer such
categories of topics.

>> Lastly, why do you guys use a hardware load balancer vs. using
zookeeper? The only reason that I could think of was to offload traffic
from the zookeeper cluster. Are there any other advantages to using a
hardware load balancer?

We have several thousand producers that talk to various Kafka clusters. At
that rate, zookeeper becomes the bottleneck. On the other hand, hardware
load balancers are designed to load balance thousands of connections across
machines, so that works better.

Thanks,
Neha
On Thu, Sep 12, 2013 at 1:03 PM, Xuyen On <[EMAIL PROTECTED]> wrote:

> Hi Neha,
>
> First I should mention that we are initially using Kafka 0.7, so I don't
> think we'll have access to the API to get metadata info from the brokers
> since this is a 0.8 feature.
>
> So my question is if we go with a hardware load balancer, do all the
> broker nodes have to treated equally? Ie: All broker nodes will have the
> same topics and number of partitions for each topic?
>
> Or is it possible to have brokers with different topics? If we go with
> this configuration how can the hardware load balancer direct the messages
> to the right broker for a given topic? We were thinking that we might be
> able to do this by having a different VIP's per topic. How do you guys do
> it?
>
> Lastly, why do you guys use a hardware load balancer vs. using zookeeper?
> The only reason that I could think of was to offload traffic from the
> zookeeper cluster. Are there any other advantages to using a hardware load
> balancer?
>
> Thanks,
>
> Xuyen
>
> -----Original Message-----
> From: Neha Narkhede [mailto:[EMAIL PROTECTED]]
> Sent: Wednesday, September 11, 2013 5:43 PM
> To: [EMAIL PROTECTED]
> Subject: Re: Hardware Loadbalancer vs zookeeper for producers
>
> Take a look at the "Producer Load balancing" section of the documentation
> - http://kafka.apache.org/08/documentation.html#theproducer
>
> Let us know how we can improve it to answer your question.
>
> Thanks,
> Neha
>
>
> On Wed, Sep 11, 2013 at 4:45 PM, Xuyen On <[EMAIL PROTECTED]> wrote:
>
> > Hi Neha,
> >
> > How do you get metadata about the cluster from the hardware load
> balancer?
> > A load balancer should direct a message to brokers that are up but I
> > don't know how you would detect for things like adding/deleting
> > topics/partitions from the cluster through the hardware load balancer.
> > Can you help me understand how you guys do this?
> >
> > Thanks,
> >
> > Xuyen
> >
> > -----Original Message-----
> > From: Neha Narkhede [mailto:[EMAIL PROTECTED]]
> > Sent: Wednesday, September 11, 2013 2:39 PM
> > To: [EMAIL PROTECTED]
> > Subject: Re: Hardware Loadbalancer vs zookeeper for producers
> >
> > Kafka 08 doesn't use zookeeper based load balancing on the producer side.
> > We rely on a hardware load balancer to be able to talk to any broker
> > and get metadata about the cluster. This happens on startup, for new