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

Switch to Threaded View
Kafka >> mail # user >> Suggestion on ZkClient usage in Kafka

Copy link to this message
Re: Suggestion on ZkClient usage in Kafka
This is being discussed in another thread -

Basically, you want zookeeper on the producer to do just one thing -
notify the change in the liveness of brokers in Kafka
cluster. In 0.8, brokers are not the entity to worry about, what we
care about are replicas for the partitions that the producer
is sending data to, in particular just the leader replica (since only
the leader can accept writes for a partition)

The producer keeps a cache of (topic, partition) -> leader-replica.
Now, if that cache is either empty or stale (due to changes
on the Kafka cluster), the next produce request will get an ACK with
an error code NotLeaderForPartition. That's when it
fires the getMetadata request that refreshes its cache. Assuming
you've configured your producer to rety (producer.num.retries)
more than once, it will succeed sending data the next time.

In other words, instead of zookeeper 'notifying' us of the changes on
the Kafka cluster, we let the producer lazily update its
cache by invoking a special API on any of the Kafka brokers. That way,
we have much fewer connections to zk, zk upgrades
are easier, so are upgrades to the producer and we also achieve the
goal of replica discovery.


On Tue, Nov 20, 2012 at 9:58 AM, Bae, Jae Hyeon <[EMAIL PROTECTED]> wrote:
> In the case that producer does not require zk.connect, how can the
> producer recognize the new brokers or brokers which went down?
> On Tue, Nov 20, 2012 at 8:31 AM, Jun Rao <[EMAIL PROTECTED]> wrote:
>> David,
>> The change in 0.8 is that instead of requiring zk.connect, we require
>> broker.list. In both cases, you typically provide a list of hosts and
>> ports. Functionality wise, they achieve the same thing, ie, the producer is
>> able to send the data to the right broker. Are you saying that zk.connect
>> is more convenient? One benefit of using broker.list is that one can
>> provide a vip as the only host. This makes it easy to add/remove brokers
>> since no producer side config needs to be changed. Changing hosts in
>> zk.connect, on the other hand, requires config changes in the client.
>> Another reason for removing zkclient in the producer is that if the client
>> GCs, it can cause churns in the producer and extra load on the zk server.
>> Since our producer can be embedded in any client, it's hard for us to
>> control the GC rate. So, removing zkclient in the producer releases the
>> potential pressure from client GC.
>> We still rely on ZK for failure detection and leader election on the broker
>> and the consumer though.
>> Thanks,
>> Jun
>> On Tue, Nov 20, 2012 at 7:54 AM, David Arthur <[EMAIL PROTECTED]> wrote:
>>> On Nov 20, 2012, at 12:23 AM, Jun Rao wrote:
>>> > Jason,
>>> >
>>> > In 0.8, producer doesn't use zkclient at all. You just need to set
>>> > broker.list.
>>> This seems like a regression in functionality. For me, one of the benefits
>>> of Kafka is only needing to know about ZooKeeper
>>> > A number of things have changed In 0.8. First, number of
>>> > partitions of a topic is global in a cluster and they don't really change
>>> > as new brokers are added. Second, a partition is assigned to multiple
>>> > brokers for replication and one of the replicas is the leader which
>>> serves
>>> > writes. When a producer starts up, it first uses the getMetadata api to
>>> > figure out the replica assignment for the relevant topic/partition. It
>>> then
>>> > issues producer request directly the broker where the leader resides. If
>>> > the leader broker goes down, the producer gets an exception and it will
>>> > re-issue the getMetadata api to obtain the information about the new
>>> leader.
>>> >
>>> > Thanks,
>>> >
>>> > Jun
>>> >
>>> > On Mon, Nov 19, 2012 at 1:29 PM, Jason Rosenberg <[EMAIL PROTECTED]>
>>> wrote:
>>> >
>>> >> Well, they do use zk though, to get the initial list of kafka nodes, and
>>> >> while zk is available, presumably they do use it to keep up with the