[MESOS-7258] Provide scheduler calls to subscribe to additional roles and unsubscribe from roles.

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Resolved
• Fix Version/s: 1.9.0
• Component/s: master, scheduler api
• Labels: multitenancy resource-management
• Epic Link: Multi-role Scalibility
• Sprint: Mesosphere Sprint 73, Mesosphere Sprint 74, Mesosphere Sprint 75, Mesosphere Sprint 76, Mesosphere Sprint 77, Mesosphere Sprint 78, Mesosphere Sprint 79, Mesosphere Sprint 2018-20, Resource Mgmt: RI-13 Sp 44, Resource Mgmt: RI13 Sp 45, Resource Mgmt: RI14 Sp 46, Resource Mgmt: RI14 Sp 47, Resource Mgmt: RI15 Sp 48
• Story Points: 13
• Assignee: Andrei Sekretenko
• Reporter: Benjamin Mahler
• Shepherd: Benjamin Mahler
• Created: 16/Mar/17 19:03
• Updated: 13/Jun/19 12:09
• Resolved: 13/Jun/19 12:09

Description

The current support for schedulers to subscribe to additional roles or unsubscribe from some of their roles requires that the scheduler obtain a new subscription with the master which invalidates the event stream.

A more lightweight mechanism would be to provide calls for the scheduler to subscribe to additional roles or unsubscribe from some roles such that the existing event stream remains open and offers to the new roles arrive on the existing event stream. E.g.

SUBSCRIBE_TO_ROLE
UNSUBSCRIBE_FROM_ROLE

One open question pertains to the terminology here, whether we would want to avoid using "subscribe" in this context. An alternative would be:

UPDATE_FRAMEWORK_INFO

Which provides a generic mechanism for a framework to perform framework info updates without obtaining a new event stream.

In addition, it would be easier to use if it returned 200 on success and an error response if invalid, etc. Rather than returning 202.

NOTE: Not specific to this issue, but we need to figure out how to allow the framework to not leak reservations, e.g. MESOS-7651.

Issue Links

• is blocked by: MESOS-9808 libprocess can deadlock on termination (cleanup() vs use() + terminate())
• is related to: MESOS-703 master fails to respect updated FrameworkInfo when the framework scheduler restarts
• relates to: MESOS-9763 Race between two re-subscriptions against an empty master.
• relates to: MESOS-9746 Consolidate the logic of sending update messages on FrameworkInfo changes.
• relates to: MESOS-9747 Make behaviour of re-subscription consistent with UPDATE_FRAMEWORK.
• relates to: MESOS-9523 Add per-framework allocatable resources matcher/filter.
• relates to: MESOS-9793 Implement UPDATE_FRAMEWORK call in V0 API for C++/Java