Rayman created SAMZA-2169:
-----------------------------
Summary: Preventing task-shuffle after task mode addition
Key: SAMZA-2169
URL: https://issues.apache.org/jira/browse/SAMZA-2169
Project: Samza
Issue Type: Bug
Reporter: Rayman
After adding task-mode, the first re-deployment of an existing job will not
have any existing task-modes to read from the coordinator stream. Due to this
the existing task-container mapping of the job does not get read.
In this case, the grouper will re-assign the tasks to containers in
lexicographic order, whereas the previous task-container mapping may have been
different than this (based on how the job changed over time – because on topic
addition a grouper tries to be as close to existing task-grouping as possible).
Due to this regrouping, task-container mapping will shuffle, causing some
containers to get new tasks, who's store need seeding from Kafka on deployment.
This can increase job re-deployment time for the first re-deploy.
Note that, after the first re-deploy the behavior becomes stable as before.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
