MLOps tools, pipelines and frameworks for flexibility and scalability in ML workflows.
Building blocks of an ML system in production:
-
Data prep:
1.1 Data Ingestion.
1.2 Data Analysis.
1.3 Data Transformation.
1.4 Data Validation. 1.5 Data Splitting. -
Model development + training at scale:
2.1 Model training.
2.2 Model Validation.
2.3 MOdel training at scale. -
Deployment and Monitoring:
3.1 Deployment.
3.2 Serving.
3.1 Monitoring.
3.2 Logging.
Pros:
- Composability.
- Portability.
- Scaling.
Amazon EKS
Pros:
- Fully managed kubernetes control plane.
- UI.
- Scheduling engine for multi-step ML workflows.
- SDK for pipeline and components manipulation.
- Pre packaged optimised deep learning docker containers are offered by EC2, SageMaker, EKS. (No need for further tuning).
Kubeflow on AWS (EKS) Set up:
- Install
kubectl
locally. - Install
AWS-CLI
. - Configure the
aws-cli
to generateconfig
andcredential
files. - Install
eksctl
. - Install the
aws-iam-authenticator
. - Create an
EKS cluster
usingeksctl
.
6.1 Export environment variables:cluster name, region, k8s version
andEC2_instance type
. - Create a cluster config file for use with eksctl and confirm its creation.
- Install Kubeflow.
- Create a kubeflow project
- Access the kubeflow UI.