Ok, the picture I have in my mind of how things work in 0.8 (from a
producer's point of view) is as follows:
1. An application program sends log messages to a producer. Each
message is provided as a key/value pair, where the key is chosen
by the application and the value is the message contents. By its
choice of key, the application may influence or control which
partition the message gets sent to.
2. The producer receives messages as key/value pairs. From talking
with zookeeper, it knows the set of available brokers and which
partitions each broker has. If the sending application provided a key
for a given message, the contents of the key may optionally
influence the producer's choice of broker and partition to send the
message to, according to some convention understood by both
application program and producer.
Is this correct?
Thanks,
Dave
On Wed, May 22, 2013 at 9:28 AM, Jun Rao <jun...@gmail.com> wrote:
> Dave,
>
> Currently, the broker expects each producer request to specify the exact
> partition id (-1 is on longer valid). The mapping from a message to a
> partition is done at the producer client. The producer can choose a random
> partition (from the existing list of partitions) or deterministically
> choose a partition based on the key.
>
> Thanks,
>
> Jun
>
>
> On Tue, May 21, 2013 at 1:12 PM, Dave Peterson <dspeter...@tagged.com>wrote:
>
>> In my case, there is a load balancer between the producers and the
>> brokers, so I want the behavior described for the Java client (null key
>> specifies "any partition"). If the Key field of each individual message
>> specifies the partition to send it to, then I don't understand the purpose
>> of the 32-bit partition identifier that precedes each message set in a
>> produce request: what if a produce request specifies "partition N" for a
>> given message set, and then each individual message in the set
>> specifies a different partition in its Key field? Also, the above-
>> mentioned partition identifier is a 32-bit integer and the Key field of
>> each individual message can contain data of arbitrary length, which
>> seems inconsistent. Is a partition identifier a 32-bit integer, or can it
>> be of arbitrary length?
>>
>> Thanks,
>> Dave
>>
>> On Tue, May 21, 2013 at 12:30 PM, Neha Narkhede <neha.narkh...@gmail.com>
>> wrote:
>> > Dave,
>> >
>> > Colin described the producer behavior of picking the partition for a
>> > message before it is sent to Kafka broker correctly. However, I'm
>> > interested in knowing your use case a little before to see why you would
>> > rather have the broker decide the partition?
>> >
>> > Thanks,
>> > Neha
>> >
>> >
>> > On Tue, May 21, 2013 at 12:05 PM, Colin Blower <cblo...@barracuda.com
>> >wrote:
>> >
>> >> The key is used by the client to decide which partition to send the
>> >> message to. By the time the client is creating the produce request, it
>> >> should be known which partition each message is being sent to. I believe
>> >> Neha described the behavior of the Java client which sends messages with
>> >> a null key to any partition.
>> >>
>> >> The key is described in past tense because of the use case for
>> >> persisting keys with messages. The key is persisted through the broker
>> >> so that a consumer knows what key was used to partition the message on
>> >> the producer side.
>> >>
>> >> I don't believe that you can have the broker decide which partition a
>> >> message goes to.
>> >>
>> >> --
>> >> Colin B.
>> >>
>> >> On 05/21/2013 11:48 AM, Dave Peterson wrote:
>> >> > I'm looking at the document entitled "A Guide to the Kafka Protocol"
>> >> > located here:
>> >> >
>> >> > https://cwiki.apache.org/KAFKA/a-guide-to-the-kafka-protocol.html
>> >> >
>> >> > It shows a produce request as containing a number of message sets,
>> which
>> >> are
>> >> > grouped first by topic and second by partition (a 32-bit integer).
>> >> > However, each
>> >> > message in a message set contains a Key field, which is described as
>> >> follows:
>> >> >
>> >> > The key is an optional message key that was used for partition
>> >> assignment.
>> >> > The key can be null.
>> >> >
>> >> > I notice the use of "was" (past tense) above. That seems to suggest
>> >> that the
>> >> > Key field was once used to specify a partition (at the granularity of
>> >> each
>> >> > individual message), but the plan for the future is to instead use the
>> >> 32-bit
>> >> > partition value preceding each message set. Is this correct? If so,
>> >> when I am
>> >> > creating a produce request for 0.8, what should I use for the 32-bit
>> >> partition
>> >> > value, and how does this relate to the Key field of each individual
>> >> message?
>> >> > Ideally, I would like to just send a produce request and let the
>> broker
>> >> choose
>> >> > the partition. How do I accomplish this in 0.8, and are there plans
>> to
>> >> change
>> >> > this after 0.8?
>> >> >
>> >> > Thanks,
>> >> > Dave
>> >> >
>> >> > On Tue, May 21, 2013 at 10:47 AM, Neha Narkhede <
>> neha.narkh...@gmail.com>
>> >> wrote:
>> >> >> No. In 0.8, if you don't specify a key for a message, it is sent to
>> any
>> >> of
>> >> >> the available partitions. In other words, the partition id is
>> selected
>> >> on
>> >> >> the partition and the server doesn't get -1 as the partition id.
>> >> >>
>> >> >> Thanks,
>> >> >> Neha
>> >> >>
>> >> >>
>> >> >> On Tue, May 21, 2013 at 9:54 AM, Dave Peterson <
>> dspeter...@tagged.com
>> >> >wrote:
>> >> >>
>> >> >>> In the version 0.8 wire format for a produce request, does a value
>> of
>> >> -1
>> >> >>> still indicate "use a random partition" as it did for 0.7?
>> >> >>>
>> >> >>> Thanks,
>> >> >>> Dave
>> >> >>>
>> >>
>> >>
>> >>
>>
