GitHub user HenryCaiHaiying opened a pull request:
https://github.com/apache/kafka/pull/1538
AFKA-3890 Kafka Streams: task assignment is not maintained on cluster
restart or rolling restart
Current task assignment in TaskAssignor is not deterministic.
During cluster restart or rolling restart, we have the same set of
participating worker nodes. But the current TaskAssignor is not able to
maintain a deterministic mapping, so about 20% partitions will be reassigned
which would cause state repopulation.
When the topology of work nodes (# of worker nodes, the TaskIds they are
carrying with) is not changed, we really just want to keep the old task
assignment.
Add the code to check whether the node topology is changing or not:
- when the prevAssignedTasks from the old clientStates is the same as the
new task list
- when there is no new node joining (its prevAssignTasks would be either
empty or conflict with some other nodes)
- when there is no node dropping out (the total of prevAssignedTasks from
other nodes would not be equal to the new task list)
When the topology is not changing, we would just use the old mapping.
I also added the code to check whether the previous assignment is balanced
(whether each node's task list is within [1/2 average -- 2 * average]), if it's
not balanced, we will still start the a new task assignment.
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/HenryCaiHaiying/kafka upstream
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/kafka/pull/1538.patch
To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:
This closes #1538
----
commit 2d6c4a18ebea38fae4e7e47a45c8d699445d2738
Author: Henry Cai <h...@pinterest.com>
Date: 2016-06-20T06:06:20Z
Add the capability to maintain the old task assignment when the node
topology is not changing
During cluster restart or rolling restart, the node topology is not
changing. But the current TaskAssignor is not able to maintain a deterministic
mapping, so about 20% partitions will be reassigned which would cause state
repopulation.
When the node topology is not changed, we really just want to keep the old
task assignment.
Add the code to check whether the node topology is changing or not:
- when the prevAssignedTasks from the old clientStates is the same as the
new task list
- when there is no new node joining (its prevAssignTasks would be either
empty or conflict with some other nodes)
- when there is no node dropping (the total of prevAssignedTasks from other
nodes would not be equal to the new task list)
When the topology is not changing, we would just use the old mapping.
----
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
