public interface ResourceManagerDriver<WorkerType extends ResourceIDRetrievable>
ResourceManagerDriveris responsible for requesting and releasing resources from/to a particular external resource manager.
|Modifier and Type||Method and Description|
The deployment specific code to deregister the application.
Initialize the deployment specific components.
Release resource to the external resource manager.
Request resource from the external resource manager.
Terminate the deployment specific components.
void initialize(ResourceEventHandler<WorkerType> resourceEventHandler, ScheduledExecutor mainThreadExecutor, Executor ioExecutor, BlockedNodeRetriever blockedNodeRetriever) throws Exception
resourceEventHandler- Handler that handles resource events.
mainThreadExecutor- Rpc main thread executor.
ioExecutor- IO executor.
blockedNodeRetriever- To retrieve all blocked nodes
void terminate() throws Exception
void deregisterApplication(ApplicationStatus finalStatus, @Nullable String optionalDiagnostics) throws Exception
This method also needs to make sure all pending containers that are not registered yet are returned.
finalStatus- The application status to report.
optionalDiagnostics- A diagnostics message or
Exception- if the application could not be deregistered.
CompletableFuture<WorkerType> requestResource(TaskExecutorProcessSpec taskExecutorProcessSpec)
This method request a new resource from the external resource manager, and tries to launch a task manager inside the allocated resource, with respect to the provided taskExecutorProcessSpec. The returned future will be completed with a worker node in the deployment specific type, or exceptionally if the allocation has failed.
Note: The returned future could be cancelled by ResourceManager. This means ResourceManager don't need this resource anymore, Driver should try to cancel this request from the external resource manager.
Note: Completion of the returned future does not necessarily mean the success of resource
allocation and task manager launching. Allocation and launching failures can still happen
after the future completion. In such cases,
will be called.
The future is guaranteed to be completed in the rpc main thread, before trying to launch
the task manager, thus before the task manager registration. It is also guaranteed that
ResourceEventHandler.onWorkerTerminated(org.apache.flink.runtime.clusterframework.types.ResourceID, java.lang.String) will not be called on the requested worker,
until the returned future is completed successfully.
taskExecutorProcessSpec- Resource specification of the requested worker.
void releaseResource(WorkerType worker)
worker- Worker node to be released, in the deployment specific type.
Copyright © 2014–2023 The Apache Software Foundation. All rights reserved.