Benjamin Mahler created MESOS-810:
-------------------------------------
Summary: New Scheduler API
Key: MESOS-810
URL: https://issues.apache.org/jira/browse/MESOS-810
Project: Mesos
Issue Type: Story
Components: c++ api, java api, python api
Reporter: Benjamin Mahler
Fix For: 1.0.0
This ticket is an effort to capture requirements and link to related tickets
for the future version of the Scheduler API. We should split these out as
needed but for now I'll just document things off the top of my head:
1. Batch status update acknowledgements. The current mechanism for
acknowledging a status update is for the call to Scheduler::statusUpdate to
return. At this point we send an acknowledgement to the slave. This simplistic
approach forces schedulers to serially persist status updates if operating in a
stateful manner, ultimately leading to scaling issues.
2. Explicit behavior. By this I mean that when a Scheduler calls something like
killTask, we'll currently implicitly drop it when we're disconnected from the
Master. Rather than implicitly doing this, we either need to deliver things
reliably or inform schedulers when their request was dropped. Returning Futures
could be very powerful here.
3. Statefulness and access to state. The scheduler driver is currently
stateless (no persistence). This means schedulers currently have to persist
state using our State abstraction or the replicated log, or through their own
persistence mechanism. Providing a stateful scheduler driver increases the
simplicity of framework schedulers substantially, and providing access to state
makes it very simple to implement a framework.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
