Benjamin Mahler created MESOS-1409:
--------------------------------------
Summary: Send status update acknowledgments through the Master.
Key: MESOS-1409
URL: https://issues.apache.org/jira/browse/MESOS-1409
Project: Mesos
Issue Type: Task
Affects Versions: 0.19.0
Reporter: Benjamin Mahler
Currently status update acknowledgments are sent directly from the scheduler
driver to the slave.
This has worked well in the sense that it removes the additional overhead of
processing acknowledgments through the master.
However, in order to provide correct reconciliation of tasks, the master needs
to hold on to terminal *unacknowledged* tasks to ensure that it can correctly
identify these until the framework receives the terminal update. Without this,
we are susceptible to MESOS-1389.
There will be additional traffic through the master from acknowledgments that
we need to be aware of. However, most of the scaling bottlenecks we have seen
in the past are in the scheduler driver's queue when we send messages faster
than the framework can process them in a persistent manner. With MESOS-1408, I
suspect that we will have a smaller number of update retries in the presence of
a heavy load on a scheduler, even with the acknowledgments going through the
master.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
