Metaflow With Kubernetes Demo

• This repository is a demo of a plugin built on Metaflow to support Kubernetes as a compute abstraction for running machine learning workflows.
• Works with Kubernetes Cluster and Minikube. Works with local and Service based Metadata Provider.
• To run the metaflow run command within container follow the instructions given in Using Metaflow with Kubernetes Cluster
• The plugin supports running metaflow within container or on local machine to orchestrate workflow on Kubernetes.
• As Metaflow is currently coupled with AWS, S3 is required as a Datastore. In Future more datastores will be supported.
• multi_step_mnist.py is a demo workflow which showcases how metaflow can be used with Kubernetes to parallely train multiple models with different hyperparameters and finally easily gather collated results.
• The flow looks like the following :

Using Metaflow with Kubernetes Cluster

• Information on Cluster and Metaflow related services setup on Kubernetes is available here.

Setting Up Environment and Running Demo

• Install Minikube or use commands in kops.sh to setup a CLUSTER on AWS. For more documentation on kops with AWS check here.
• Plugin Requires AWS KEYS to be set in environment variables.
• Replace in AWS_ACCESS_KEY_ID,AWS_SECRET_ACCESS_KEY,AWS_SESSION_TOKEN in setup.sh
• Add the S3 Bucket location in setup.sh to METAFLOW_DATASTORE_SYSROOT_S3 environment variable.
• Run sh setup.sh to use Minikube and run the flow with plugin repo on Kubernetes.
• If you are not using Minikube, a Kube config is required.

Using The Plugin

• Usage is very similar to @batch decorator.
• on top of any @step add the @kube decorator or use --with kube:cpu=2,memory=4000,image=python:3.7 in the CLI args.
• To directly deploy the entire runtime into Kubernetes as a job, using the kube-deploy run command:
  • python multi_step_mnist.py --with kube:cpu=3.2,memory=4000,image=tensorflow/tensorflow:latest-py3 kube-deploy run --num_training_examples 1000 --dont-exit
  • --dont-exit will follow log trail from the job. Otherwise the workflow will be deployed as a job on Kubernetes which will destroy itself once it ends.
  • Directly deploy to kubernetes only works with Service based Metaprovider
  • Good practice before directly moving to kube-deploy would be:
    • Local tests : python multi_step_mnist.py run --num_training_examples 1000 : With or without Conda.
    • Dry run with python multi_step_mnist.py --with kube:cpu=3.2,memory=4000,image=tensorflow/tensorflow:latest-py3 run --num_training_examples 1000
    • On successful dry run : python multi_step_mnist.py --with kube:cpu=3.2,memory=4000,image=tensorflow/tensorflow:latest-py3 kube-deploy run --num_training_examples 50000 : Run Larger Dataset.

Running with Conda

• To run with Conda it will need 'python-kubernetes':'10.0.1' in the libraries argument to @conda_base step decorator
• Use image=python:3.6 when running with Conda in --with kube:. Ideally that should be the python version used/mentioned in conda.
• Direct deploy to kubernetes with Conda environment is supported
  • python multi_step_mnist.py --with kube:cpu=3.2,memory=4000,image=python:3.6 --environment=conda kube-deploy run --num_training_examples 1000 --dont-exit
  • Ensure to use image=python:<conda_python_version>

Running on GPU Clusters.

• Instructions for GPU Based setup for Metaflow on Kubernetes can be found here
• The deployment works for Cuda 10.2,9.1.
• Results of GPU Run on experiments_analytics_gpumnist.ipynb

CLI Operations Available with Kube:

• python multi_step_mnist.py kube list : Show the currently running jobs of flow.
• python multi_step_mnist.py kube kill : Kills all jobs on Kube. Any Metaflow Runtime accessing those jobs will be gracefully exited.
• python multi_step_mnist.py kube-deploy run : Will run the Metaflow Runtime inside a container on kubernetes cluster. Needs metadata service to work.
• python multi_step_mnist.py kube-deploy list : It will list any running deployment of the current flow on Kubernetes.

Seeing Results Post Completion

• check experiments_analytics.ipynb
• Before running notebook use %set_env in notebook to set env vars for AWS Keys.

Current Constraints

• Supports S3 based Datastore. Wont work without S3 as datastore.
• Current Has Been Tested with GPU's. Current Deployment Automation scripts Supports Cuda v9.1.
• Needs METAFLOW_KUBE_CONFIG_PATH and METAFLOW_KUBE_NAMESPACE env vars for kubernetes config file and Namespace . Takes ~/.kube/config and default as defaults
• Requires AWS Secrets in ENV Vars for Deploying PODS. That Needs fixing.
• Supports Conda Decorators. Current setup.sh uses tensorflow/tensorflow image.
• To use conda Add the following :
  • @conda_base(python=get_python_version(),libraries={'numpy':'1.18.1','tensorflow':'1.4.0','python-kubernetes':'10.0.1'}) above the class definition to make it work with conda.
  • TO run with conda change the line in setup.sh to following :
    • .env/bin/python multi_step_mnist.py --environment=conda --with kube:cpu=2,memory=4000,image=python:3.7 run --num_training_examples 2500
    • Please also ensure that Conda is in the PATH env variable.

Installing Plugin Repo

• pip install https://github.com/valayDave/metaflow/archive/kube_cpu_stable.zip

Plugin Fork Repo

https://github.com/valayDave/metaflow/tree/kube_cpu_stable