[
https://issues.apache.org/jira/browse/KAFKA-4402?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jun Yao updated KAFKA-4402:
---------------------------
Description:
>From this code comments, it is said that Kafka client Producer's
>DefaultPartitioner will do round robin if "no partition or key is present",
https://github.com/apache/kafka/blob/41e676d29587042994a72baa5000a8861a075c8c/clients/src/main/java/org/apache/kafka/clients/producer/internals/DefaultPartitioner.java#L34
from the code it looks trying to do round robin as well, as it maintained a
counter and try to increase it every time and then will decide which partition
to go to;
However the issue here is the counter is a global counter that is shared by all
the topics, so it is actually not round robin per topic and sometimes caused
unbalanced routing among different partitions.
Although we can pass a custom implementation of interface
"org.apache.kafka.clients.producer.Partitioner", it might be still good to make
the default implementation true round robin as comment.
> Kafka Producer's DefaultPartitioner is actually not round robin as said in
> the code comments "If no partition or key is present choose a partition in a
> round-robin fashion"
> ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: KAFKA-4402
> URL: https://issues.apache.org/jira/browse/KAFKA-4402
> Project: Kafka
> Issue Type: Improvement
> Reporter: Jun Yao
> Priority: Minor
>
> From this code comments, it is said that Kafka client Producer's
> DefaultPartitioner will do round robin if "no partition or key is present",
> https://github.com/apache/kafka/blob/41e676d29587042994a72baa5000a8861a075c8c/clients/src/main/java/org/apache/kafka/clients/producer/internals/DefaultPartitioner.java#L34
> from the code it looks trying to do round robin as well, as it maintained a
> counter and try to increase it every time and then will decide which
> partition to go to;
> However the issue here is the counter is a global counter that is shared by
> all the topics, so it is actually not round robin per topic and sometimes
> caused unbalanced routing among different partitions.
> Although we can pass a custom implementation of interface
> "org.apache.kafka.clients.producer.Partitioner", it might be still good to
> make the default implementation true round robin as comment.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
