As one standalone component of Microsoft OpenPAI, FrameworkController (FC) is built to orchestrate all kinds of applications on Kubernetes by a single controller, especially for DeepLearning applications.

These kinds of applications include but not limited to:

• Stateless and Stateful Service:
  • DeepLearning Serving: TensorFlow Serving, etc.
  • Big Data Serving: HDFS, HBase, Kafka, Etcd, Nginx, etc.
• Stateless and Stateful Batch:
  • DeepLearning AllReduce Training: TensorFlow MultiWorkerMirrored Training, Horovod Training, etc.
  • DeepLearning Elastic Training without Server: PyTorch Elastic Training with whole cluster shared etcd, etc.
  • DeepLearning Batch/Offline Inference: PyTorch Inference, etc.
  • Automated Machine Learning: NNI, etc.
  • Big Data Batch Processing: Standalone Spark, KD-Tree Building, etc.
• Any combination of above applications:
  • DeepLearning ParameterServer Training: TensorFlow ParameterServer Training, etc.
  • DeepLearning Interactive Training: TensorFlow with Jupyter Notebook, etc.
  • DeepLearning Elastic Training with Server: PyTorch Elastic Training with per-application dedicated etcd, etc.
  • DeepLearning Streaming/Online Inference: TensorFlow Inference with Streaming I/O, etc.
  • DeepLearning Incremental/Online Training: TensorFlow Training with Streaming I/O, etc.
  • Big Data Stream Processing: Standalone Flink, etc.

In the open source community, there are so many specialized Kubernetes Pod controllers which are built for a specific kind of application, such as Kubernetes StatefulSet Controller, Kubernetes Job Controller, KubeFlow TensorFlow Operator, KubeFlow PyTorch Operator. However, no one is built for all kinds of applications and combination of the existing ones still cannot support some kinds of applications. So, we have to learn, use, develop, deploy and maintain so many Pod controllers.

Build a General-Purpose Kubernetes Pod Controller: FrameworkController.

And then we can get below benefits from it:

• Support Kubernetes official unsupported applications:
  • Stateful Batch with Service applications, like TensorFlow ParameterServer Training on FC.
  • ScaleUp/ScaleDown Tolerable Stateful Batch applications, like PyTorch Elastic Training on FC.
• Only need to learn, use, develop, deploy and maintain a single controller
• All kinds of applications can leverage almost all provided features and guarantees
• All kinds of applications can be used through the same interface with a unified experience
• If really required, only need to build specialized controllers on top of it, instead of building from scratch:
  • The similar practice is also adopted by Kubernetes official controllers, such as the Kubernetes Deployment Controller is built on top of the Kubernetes ReplicaSet Controller.

A Framework represents an application with a set of Tasks:

1. Executed by Kubernetes Pod
2. Partitioned to different heterogeneous TaskRoles which share the same lifecycle
3. Ordered in the same homogeneous TaskRole by TaskIndex
4. With consistent identity {FrameworkName}-{TaskRoleName}-{TaskIndex} as PodName
5. With fine grained ExecutionType to Start/Stop the whole Framework
6. With fine grained RetryPolicy for each Task and the whole Framework
7. With fine grained FrameworkAttemptCompletionPolicy for each TaskRole
8. With PodGracefulDeletionTimeoutSec for each Task to tune Consistency vs Availability
9. With fine grained Status for each TaskAttempt/Task, each TaskRole and the whole FrameworkAttempt/Framework

1. Highly generalized as it is built for all kinds of applications
2. Light-weight as it is only responsible for Pod orchestration
3. Well-defined Framework Consistency vs Availability, State Machine and Failure Model
4. Tolerate Pod/ConfigMap unexpected deletion, Node/Network/FrameworkController/Kubernetes failure
5. Support to specify how to classify and summarize Pod failures
6. Support to ScaleUp/ScaleDown Framework with Strong Safety Guarantee
7. Support to expose Framework and Pod history snapshots to external systems
8. Easy to leverage FrameworkBarrier to achieve light-weight Gang Execution and Service Discovery
9. Easy to leverage HiveDScheduler to achieve GPU Topology-Aware, Multi-Tenant, Priority and Gang Scheduling
10. Compatible with other Kubernetes features, such as Kubernetes Service, Gpu Scheduling, Volume, Logging
11. Idiomatic with Kubernetes official controllers, such as Pod Spec
12. Aligned with Kubernetes Controller Design Guidelines and API Conventions

1. A Kubernetes cluster, v1.14.2 or above, on-cloud or on-premise.

1. Run Controller
2. Submit Framework

1. User Manual
2. Known Issue and Upcoming Feature
3. FAQ
4. Release Note

• DockerHub

A specialized wrapper can be built on top of FrameworkController to optimize for a specific kind of application:

• Microsoft OpenPAI Controller Wrapper (Job RestServer): A wrapper client optimized for AI applications
• Microsoft DLWorkspace Controller Wrapper (Job Manager): A wrapper client optimized for AI applications
• Microsoft NNI Controller Wrapper (TrainingService): A wrapper client optimized for AutoML applications

FrameworkController can directly leverage many Kubernetes Schedulers and among them we recommend these best fits:

• Kubernetes Default Scheduler: A General-Purpose Kubernetes Scheduler
• HiveDScheduler: A Kubernetes Scheduler Extender optimized for AI applications

• YARN FrameworkLauncher: Similar offering on Apache YARN

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.