Maximilian Michels created FLINK-3543:
-----------------------------------------
Summary: Introduce ResourceManager component
Key: FLINK-3543
URL: https://issues.apache.org/jira/browse/FLINK-3543
Project: Flink
Issue Type: New Feature
Components: ResourceManager, JobManager, TaskManager
Affects Versions: 1.1.0
Reporter: Maximilian Michels
Assignee: Maximilian Michels
Fix For: 1.1.0
So far the JobManager has been the central instance which is responsible for
resource management and allocation.
While thinking about how to integrate Mesos support in Flink, people from the
Flink community realized that it would be nice to delegate resource allocation
to a dedicated process. This process may run independently of the JobManager
which is a requirement for proper integration of cluster allocation frameworks
like Mesos.
This has led to the idea of creating a new component called the
{{ResourceManager}}. Its task is to allocate and maintain resources requested
by the {{JobManager}}. The ResourceManager has a very abstract notion of
resources.
Initially, we thought we could make the ResourceManager deal with resource
allocation and the registration/supervision of the TaskManagers. However, this
approach proved to add unnecessary complexity to the runtime. Registration
state of TaskManagers had to be kept in sync at both the JobManager and the
ResourceManager.
That's why [~StephanEwen] and me changed the ResourceManager's role to simply
deal with the resource acquisition. The TaskManagers still register with the
JobManager which informs the ResourceManager about the successful registration
of a TaskManager. The ResourceManager may inform the JobManager of failed
TaskManagers. Due to the insight which the ResourceManager has in the resource
health, it may detect failed TaskManagers much earlier than the heartbeat-based
monitoring of the JobManager.
At this stage, the ResourceManager is an optional component. That means the
JobManager doesn't depend on the ResourceManager as long as it has enough
resources to perform the computation. All bookkeeping is performed by the
JobManager. When the ResourceManager connects to the JobManager, it receives
the current resources, i.e. task manager instances, and allocates more
containers if necessary. The JobManager adjusts the number of containers
through the {{SetWorkerPoolSize}} method.
In standalone mode, the ResourceManager may be deactivated or simply use the
StandaloneResourceManager which does practically nothing because we don't need
to allocate resources in standalone mode.
In YARN mode, the ResourceManager takes care of communicating with the Yarn
resource manager. When containers fail, it informs the JobManager and tries to
allocate new containers. The ResourceManager runs as an actor within the same
actor system as the JobManager. It could, however, also run independently. The
independent mode would be the default behavior for Mesos where the framework
master is expected to just deal with resource allocation.
The attached figures depict the message flow between ResourceManager,
JobManager, and TaskManager.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
