THIS IS A TEST INSTANCE. ALL YOUR CHANGES WILL BE LOST!!!!
...
In the new producer you can pass in a key or hard code the partition as part of ProducerRecord.
Internally we are using a class
This class uses the specified partition if there is one; uses a hash of the key if there isn't a partition but there is a key; and simply chooses a partition round robin if there is neither a partition nor a key.
However there are several partitioning strategies that could be useful that we don't support out of the box.
An example would be having each producer periodically choose a random partition. This tends to be the most efficient since all data goes to one server and uses the fewest TCP connections, however it only produces good load balancing if there are many producers.
Of course a user can do this now by just setting the partition manually, but that is a bit inconvenient if you need to do that across a bunch of apps since each will need to remember to set the partition every time.
The idea would be to expose a configuration to set the partitioner implementation like
partitioner.class=org.apache.kafka.producer.DefaultPartitioner
This would default to the existing partitioner implementation.
Public Interfaces
| Code Block | ||
|---|---|---|
| ||
package org.apache.kafka.clients.producer; |
...
public interface Partitioner { |
...
/** |
...
* Compute the partition for the given record. |
...
*
* @param topic The topic name
* @param key The key to partition on (or null if no key)
* @param partition The partition to use (or null if none)
* @param cluster The current cluster metadata
*/
public int partition(String topic, byte[] key, Integer partition, Cluster cluster);
} |
| Code Block | ||
|---|---|---|
| ||
package |
}
...
org.apache.kafka.clients.producer; |
...
public class DefaultPartitioner implements partitioner { } |
}
Proposed Changes
- Add a new partitioner interface .
- Move existing partitioner code to DefaultPartitioner .
- Introduce new producer config called "partitioner.class" . By default it points to org.apache.kafka.clients.producer.DefaultPartitioner
...
- It will not impact any of existing clients. Even when they upgrade to they new version, they can continue to use their existing client code without the need to add partitioner.class config.
...