flink-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ufuk Celebi (JIRA)" <j...@apache.org>
Subject [jira] [Created] (FLINK-1672) Refactor task registration/unregistration
Date Tue, 10 Mar 2015 10:57:38 GMT
Ufuk Celebi created FLINK-1672:

             Summary: Refactor task registration/unregistration
                 Key: FLINK-1672
                 URL: https://issues.apache.org/jira/browse/FLINK-1672
             Project: Flink
          Issue Type: Improvement
          Components: Distributed Runtime
            Reporter: Ufuk Celebi

h4. Current control flow for task registrations

# JM submits a TaskDeploymentDescriptor to a TM
## TM registers the required JAR files with the LibraryCacheManager and returns the user code
class loader
## TM creates a Task instance and registers the task in the runningTasks map
## TM creates a TaskInputSplitProvider
## TM creates a RuntimeEnvironment and sets it as the environment for the task
## TM registers the task with the network environment
## TM sends async msg to profiler to monitor tasks
## TM creates temporary files in file cache
## TM tries to start the task

If any operation >= 1.2 fails:
* TM calls task.failExternally()
* TM removes temporary files from file cache
* TM unregisters the task from the network environment
* TM sends async msg to profiler to unmonitor tasks
* TM calls unregisterMemoryManager on task

If 1.1 fails, only unregister from LibraryCacheManager.

h4. RuntimeEnvironment, Task, TaskManager separation

The RuntimeEnvironment has references to certain components of the task manager like memory
manager, which are accecssed from the task. Furthermore it implements Runnable, and creates
the executing task Thread. The Task instance essentially wraps the RuntimeEnvironment and
allows asynchronous state management of the task (RUNNING, FINISHED, etc.).

The way that the state updates affect the task is not that obvious: state changes trigger
messages to the TM, which for final states further trigger a msg to unregister the task. The
way that tasks are unregistered again depends on the state of the task.


I would propose to refactor this to make the way the state handling/registration/unregistration
is handled is more transparent.

This message was sent by Atlassian JIRA

View raw message