Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/Microsoft/frameworkcontroller

General-Purpose Kubernetes Pod Controller
https://github.com/Microsoft/frameworkcontroller

container-management container-orchestration containers frameworklauncher go kubernetes kubernetes-controller tensorflow

Last synced: about 2 months ago
JSON representation

General-Purpose Kubernetes Pod Controller

Awesome Lists containing this project

README 

        
**_Turn it into maintenance mode, read-only, and only update for critical fixes._**



# Microsoft OpenPAI FrameworkController



[![Build Status](https://github.com/microsoft/frameworkcontroller/workflows/build/badge.svg?branch=master&event=push)](https://github.com/microsoft/frameworkcontroller/actions?query=workflow%3Abuild+branch%3Amaster+event%3Apush)

[![Latest Release](https://img.shields.io/github/release/microsoft/frameworkcontroller.svg)](https://github.com/microsoft/frameworkcontroller/releases/latest)

[![Docker Pulls](https://img.shields.io/docker/pulls/frameworkcontroller/frameworkcontroller.svg)](https://hub.docker.com/u/frameworkcontroller)

[![License](https://img.shields.io/github/license/microsoft/frameworkcontroller.svg)](https://github.com/microsoft/frameworkcontroller/blob/master/LICENSE)



As one standalone component of [Microsoft OpenPAI](https://github.com/microsoft/pai), FrameworkController (FC) is built to orchestrate all kinds of applications on [Kubernetes](https://kubernetes.io) 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](https://www.tensorflow.org/tfx/guide/serving), etc.

  - Big Data Serving: HDFS, HBase, Kafka, Etcd, Nginx, etc.

- __Stateless and Stateful Batch__:

  - DeepLearning AllReduce Training: [TensorFlow MultiWorkerMirrored Training](https://www.tensorflow.org/guide/distributed_training#multiworkermirroredstrategy), [Horovod Training](https://horovod.readthedocs.io), etc.

  - DeepLearning Elastic Training without Server: [PyTorch Elastic Training with whole cluster shared etcd](https://pytorch.org/elastic), etc.

  - DeepLearning Batch/Offline Inference: [PyTorch Inference](https://pytorch.org/tutorials/recipes/recipes/saving_and_loading_models_for_inference.html), etc.

  - Automated Machine Learning: [NNI](https://nni.readthedocs.io), etc.

  - Big Data Batch Processing: [Standalone Spark](http://spark.apache.org/docs/latest/spark-standalone.html), KD-Tree Building, etc.

- __Any combination of above applications__:

  - DeepLearning ParameterServer Training: [TensorFlow ParameterServer Training](https://www.tensorflow.org/guide/distributed_training#parameterserverstrategy), etc.

  - DeepLearning Interactive Training: [TensorFlow with Jupyter Notebook](https://www.tensorflow.org/tensorboard/tensorboard_in_notebooks), etc.

  - DeepLearning Elastic Training with Server: [PyTorch Elastic Training with per-application dedicated etcd](https://pytorch.org/elastic), etc.

  - DeepLearning Streaming/Online Inference: [TensorFlow Inference with Streaming I/O](https://www.tensorflow.org/io), etc.

  - DeepLearning Incremental/Online Training: [TensorFlow Training with Streaming I/O](https://www.tensorflow.org/io), etc.

  - Big Data Stream Processing: [Standalone Flink](https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/cluster_setup.html), etc.



## Why Need It

### Problem

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](https://kubernetes.io/docs/concepts/workloads/controllers/statefulset), [Kubernetes Job Controller](https://kubernetes.io/docs/concepts/workloads/controllers/job), [KubeFlow TensorFlow Operator](https://github.com/kubeflow/tf-operator), [KubeFlow PyTorch Operator](https://github.com/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.



### Solution

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](example/framework/basic/batchwithservicesucceeded.yaml) applications, like [TensorFlow ParameterServer Training on FC](example/framework/scenario/tensorflow/ps).

  - [ScaleUp/ScaleDown Tolerable Stateful Batch](doc/user-manual.md#FrameworkRescale) applications, like [PyTorch Elastic Training on FC](example/framework/scenario/pytorch/elastic).

- __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](https://kubernetes.io/docs/concepts/workloads/controllers/deployment) is built on top of the [Kubernetes ReplicaSet Controller](https://kubernetes.io/docs/concepts/workloads/controllers/replicaset).



## Architecture



  Architecture




## Feature

### Framework Feature

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](doc/user-manual.md#FrameworkExecutionType) to Start/Stop the whole Framework

6. With fine grained [RetryPolicy](doc/user-manual.md#RetryPolicy) for each Task and the whole Framework

7. With fine grained [FrameworkAttemptCompletionPolicy](doc/user-manual.md#FrameworkAttemptCompletionPolicy) for each TaskRole

8. With PodGracefulDeletionTimeoutSec for each Task to [tune Consistency vs Availability](doc/user-manual.md#FrameworkConsistencyAvailability)

9. With fine grained [Status](pkg/apis/frameworkcontroller/v1/types.go) for each TaskAttempt/Task, each TaskRole and the whole FrameworkAttempt/Framework



### Controller Feature

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](doc/user-manual.md#FrameworkConsistencyAvailability), [State Machine](doc/user-manual.md#FrameworkTaskStateMachine) and [Failure Model](doc/user-manual.md#CompletionStatus)

4. Tolerate Pod/ConfigMap unexpected deletion, Node/Network/FrameworkController/Kubernetes failure

5. Support to specify how to [classify and summarize Pod failures](doc/user-manual.md#PodFailureClassification)

6. Support to [ScaleUp/ScaleDown Framework with Strong Safety Guarantee](doc/user-manual.md#FrameworkRescale)

7. Support to expose [Framework and Pod history snapshots](doc/user-manual.md#FrameworkPodHistory) to external systems

8. Easy to leverage [FrameworkBarrier](doc/user-manual.md#FrameworkBarrier) to achieve light-weight Gang Execution and Service Discovery

9. Easy to leverage [HiveDScheduler](doc/user-manual.md#HiveDScheduler) to achieve GPU Topology-Aware, Multi-Tenant, Priority and Gang Scheduling

10. Compatible with other Kubernetes features, such as Kubernetes [Service](https://kubernetes.io/docs/concepts/services-networking/service), [Gpu Scheduling](https://kubernetes.io/docs/tasks/manage-gpus/scheduling-gpus), [Volume](https://kubernetes.io/docs/concepts/storage/volumes), [Logging](https://kubernetes.io/docs/concepts/cluster-administration/logging)

11. Idiomatic with Kubernetes official controllers, such as [Pod Spec](https://kubernetes.io/docs/concepts/workloads/pods/pod-overview/#pod-templates)

12. Aligned with Kubernetes [Controller Design Guidelines](https://github.com/kubernetes/community/blob/f0dd87ad477e1e91c53866902adf7832c32ce543/contributors/devel/sig-api-machinery/controllers.md) and [API Conventions](https://github.com/kubernetes/community/blob/a2cdce51a0bbbc214f0e8813e0a877176ad3b6c9/contributors/devel/sig-architecture/api-conventions.md)



## Prerequisite

1. A Kubernetes cluster, v1.16.15 or above, on-cloud or on-premise.



## Quick Start

1. [Run Controller](example/run)

2. [Submit Framework](example/framework)



## Doc

1. [Deep Dive Slides](doc/deep-dive.pptx)

2. [User Manual](doc/user-manual.md)

3. [Known Issue and Upcoming Feature](doc/known-issue-and-upcoming-feature.md)

4. FAQ

5. Release Note



## Official Image

* [DockerHub](https://hub.docker.com/u/frameworkcontroller)



## Related Project

### Third Party Controller Wrapper

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

* [Microsoft OpenPAI Controller Wrapper (Job RestServer)](https://github.com/microsoft/pai/tree/master/src/rest-server): A wrapper client optimized for AI applications

* [Microsoft AzureML Kubernetes Compute Controller Wrapper](https://k8s-wiki.azureml.com/overview/index.html): A wrapper client optimized for AI applications: *AzureML Kubernetes Compute or ITP (Integrated Training Platform) is built for both first party and third party users, and will be eventually leveraged by [AML (Azure Machine Learning)](https://azure.microsoft.com/en-us/services/machine-learning)*

* [Microsoft DLWorkspace Controller Wrapper (Job Manager)](https://github.com/microsoft/DLWorkspace/blob/914f347d18e852bc6a6d3e86fe25ac040a3f78f9/src/ClusterManager/job_manager.py): A wrapper client optimized for AI applications

* [Microsoft NNI Controller Wrapper (TrainingService)](https://github.com/microsoft/nni/blob/master/docs/en_US/TrainingService/FrameworkControllerMode.md): A wrapper client optimized for AutoML applications



### Recommended Kubernetes Scheduler

FrameworkController can directly leverage many [Kubernetes Schedulers](https://kubernetes.io/docs/tasks/administer-cluster/configure-multiple-schedulers) and among them we recommend these best fits:

* [Kubernetes Default Scheduler](https://kubernetes.io/docs/concepts/scheduling/kube-scheduler/#kube-scheduler): A General-Purpose Kubernetes Scheduler

* [HiveDScheduler](doc/user-manual.md#HiveDScheduler): A Kubernetes Scheduler Extender optimized for AI applications



### Similar Offering On Other Cluster Manager

* [YARN FrameworkLauncher](https://github.com/microsoft/pai/blob/master/subprojects/frameworklauncher/yarn): Similar offering on [Apache YARN](http://hadoop.apache.org)



## Contributing

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](https://opensource.microsoft.com/codeofconduct/).

For more information see the [Code of Conduct FAQ](https://opensource.microsoft.com/codeofconduct/faq/) or

contact [[email protected]](mailto:[email protected]) with any additional questions or comments.