As a consumer, how can I get partition count for a topic

[winbatch]

Let's say I have a known algorithm for choosing the partition in my publisher. However, that algorithm takes the partition count into consideration via the partitioner_callback. How can I get the partition count in a consumer so I can apply a similar algorithm?

[edenhill]

At what time do you want the partition count?

• before calling consume_start()
• updated periodically (e.g., every 10s)
• for each consumed messages

[winbatch]

In this particular case, before calling consume start

On Friday, March 21, 2014, Magnus Edenhill notifications@github.com wrote:

At what time do you want the partition count?

• before calling consume_start()
• updated periodically (e.g., every 10s)
• for each consumed messages

Reply to this email directly or view it on GitHubhttps://github.com//issues/94#issuecomment-38339854
.

[edenhill]

Ok, your program would have to wait for this information to be available (rdkafka needs to connect to brokers, query metadata, etc). I see two alternatives here:

• async approach: you register a callback for topic metadata updates which will be called when information regarding a topic changes (partition count, leader change, ..)
• sync approach: will block until the metadata is available, possibly with a timeout argument - this is sketchy since timeout values are arbitrary.

The async approach is preferred and the application can provide its own sync version on top of it if necessary.

Comment?

[edenhill]

This is related to issue #90

[winbatch]

that's fine. In this particular use case, I just want it at program
startup. Basically, I'm implementing a simple kafka 'grep' so that someone
gives me a topic and I find their message. One of the things they pass in
is the key, and I'll call my algorithm for determining which of the
partitions to look on. (And since my algorithm requires the # of
partitions I need this info)

On Fri, Mar 21, 2014 at 10:19 PM, Magnus Edenhill
notifications@github.comwrote:

Ok, your program would have to wait for this information to be available
(rdkafka needs to connect to brokers, query metadata, etc). I see two
alternatives here:

• async approach: you register a callback for topic metadata updates
  which will be called when information regarding a topic changes (partition
  count, leader change, ..)
• sync approach: will block until the metadata is available, possibly
  with a timeout argument - this is sketchy since timeout values are
  arbitrary.

The async approach is preferred and the application can provide its own
sync version on top of it if necessary.

Comment?

Reply to this email directly or view it on GitHubhttps://github.com//issues/94#issuecomment-38340524
.

[edenhill]

Okay, so then your program would look something like:

int the_part_cnt = -1;
topic_metadata_callback(..) {
    the_part_cnt = good_intel.part_cnt;
}

main() {

   rd_kafka_new();
   ... register callbacks bla bla...

  wait_for(the_part_cnt != -1);

  part = partitioner(key, the_part_cnt);

  consume_start(,..part);

   ...consume messages...

 ...decommission...
}

Right?

[winbatch]

looks reasonable

On Fri, Mar 21, 2014 at 10:30 PM, Magnus Edenhill
notifications@github.comwrote:

Okay, so then your program would look something like:

int the_part_cnt = -1;
topic_metadata_callback(..) {
the_part_cnt = good_intel.part_cnt;
}

main() {

rd_kafka_new();
... register callbacks bla bla...

wait_for(the_part_cnt != -1);

part = partitioner(key, the_part_cnt);

consume_start(,..part);

...consume messages...

...decommission...
}

Right?

Reply to this email directly or view it on GitHubhttps://github.com//issues/94#issuecomment-38340752
.

[edenhill]

The topic metadata is a bit complex, it will be formatted something like this:

void your_topic_info_callback(rd_kafka_topic_t *rkt, const struct *TopicInfo, void *rkt_opaque);

struct TopicInfo {
    int partition_cnt;
    struct PartitionInfo[partition_cnt];
};

struct PartitionInfo {
   int32_t partition_id;
   int32_t leader;
   int replica_cnt;
   int32_t replicas[replica_cnt];
   int isr_cnt;
   int32_t isrs[isr_cnt];
};

[edenhill]

Have a look in rdkafka_example.c (look for metadata) for how to use the new Metadata API.

This also adds a handy -L switch rdkafka_example:

$ examples/rdkafka_example -L -t onepart
Metadata for onepart (from broker -1):
 1 brokers:
  broker 3 at bib.local:9093
 1 topics:
  topic "onepart" with 2 partitions:
    partition 0, leader 3, replicas: 3, isrs: 3
    partition 1, leader 3, replicas: 3, isrs: 3
$ examples/rdkafka_example -L | head -30
Metadata for all topics (from broker -1):
 3 brokers:
  broker 1 at bib.local:9091
  broker 2 at bib.local:9092
  broker 3 at bib.local:9093
 14 topics:
  topic "z7" with 2 partitions:
    partition 0, leader 1, replicas: 1, isrs: 1
    partition 1, leader 2, replicas: 2, isrs: 2
  topic "rdkafkatest1_auto_49f744a4327b1b1e" with 2 partitions:
    partition 0, leader 3, replicas: 3, isrs: 3
    partition 1, leader 1, replicas: 1, isrs: 1
  topic "rdkafkatest1_auto_e02f58f2c581cba" with 2 partitions:
    partition 0, leader 3, replicas: 3, isrs: 3
    partition 1, leader 1, replicas: 1, isrs: 1
  topic "aba" with 2 partitions:
    partition 0, leader 3, replicas: 3, isrs: 3
    partition 1, leader 1, replicas: 1, isrs: 1
  topic "aba3" with 2 partitions:
    partition 0, leader 1, replicas: 1, isrs: 1
    partition 1, leader 2, replicas: 2, isrs: 2
  topic "z" with 2 partitions:
    partition 0, leader 1, replicas: 1, isrs: 1
    partition 1, leader 2, replicas: 2, isrs: 2
  topic "onepartx" with 2 partitions:
    partition 0, leader 1, replicas: 1, isrs: 1
    partition 1, leader 2, replicas: 2, isrs: 2
  topic "rdkafkatest2" with 2 partitions:
    partition 0, leader 1, replicas: 1, isrs: 1
    partition 1, leader 2, replicas: 2, isrs: 2
...

[edenhill]

Please verify this on your end aswell.