[jira] [Created] (KAFKA-4553) Connect's round robin assignment produces undesirable distribution of connectors/tasks

["Ewen Cheslack-Postava (JIRA)" <ji...@apache.org>]

Ewen Cheslack-Postava created KAFKA-4553:
--------------------------------------------

             Summary: Connect's round robin assignment produces undesirable distribution of connectors/tasks
                 Key: KAFKA-4553
                 URL: https://issues.apache.org/jira/browse/KAFKA-4553
             Project: Kafka
          Issue Type: Bug
          Components: KafkaConnect
    Affects Versions: 0.10.1.0
            Reporter: Ewen Cheslack-Postava
            Assignee: Ewen Cheslack-Postava


Currently the round robin assignment in Connect looks something like this:

foreach connector {
  assign connector to next worker
  for each task in connector {
    assign task to next member
  }
}

For the most part we assume that connectors and tasks are effectively equivalent units of work, but this is actually rarely the case. Connectors are usually much lighterweight as they are just monitoring for changes in the source/sink system and tasks are doing the heavy lifting. The way we are currently doing round robin assignment then causes uneven distributions of work in some cases that are not too uncommon.

In particular, it gets bad if there are an even number of workers and connectors that generate only a single task since this results in the even #'d workers always getting assigned connectors and odd workers always getting assigned tasks. An extreme case of this is when users start distributed mode clusters with just a couple of workers to get started and deploy multiple single-task connectors (e.g. CDC connectors like Debezium would be a common example). All the connectors end up on one worker, all the tasks end up on the other, and the second worker becomes overloaded.

Although the ideal solution to this problem is to have a better idea of how much load each connector/task will generate, I don't think we want to get into the business of full-on cluster resource management. An alternative which I think avoids this common pitfall without the risk of hitting another common bad case is to change the algorithm to assign all the connectors first, then all the tasks, i.e.

foreach connector {
  assign connector to next worker
}
foreach connector {
  for each task in connector {
    assign task to next worker
  }
}




--
This message was sent by Atlassian JIRA
(v6.3.4#6332)