qiaoxingli/amazon-sagemaker-mlops-workshop

Machine Learning Ops Workshop with SageMaker: lab guides and materials.

Amazon Sagemaker MLops (with classic CI/CD tools) Workshop

Machine Learning Ops Workshop with SageMaker and CodePipeline: lab guides and materials.

Introduction

Data Scientists and ML developers need more than a Jupyter notebook to create a ML model, to test it, to put it into production and to integrate it with a portal and/or a basic web/mobile application, in a reliable and flexible way.

There are two basic questions that you should consider when you start developing a ML model for a real Business Case:

1. How long would it take your organization to deploy a change that involves a single line of code?
2. Can you do this on a repeatable, reliable basis?

So, if you're not happy with the answers you have, MLOps is a concept that can help you: a) to create or improve the organization culture for CI/CD applied to ML; b) to create an automated infrastructure that will support your processes.

In this workshop you'll see how to create/operate an automated ML pipeline using a traditional CI/CD tool, called CodePipeline, to orchestrate the ML workflow. During the exercises you'll see how to create a Docker container from scratch with your own algorithm, start a training/deployment job by just copying a .zip file to an S3 repo, run A/B tests and more. This is a reference architecture that can be used as an inspiration to create your own solution.

Amazon Sagemaker, a service that supports the whole pipeline of a ML Model development lifecycle, is the heart of this solution. Around it, you can add several different services as the AWS Code* for creating an automated pipeline, building your docker images, train/test/deploy/integrate your models, etc.

Here you can find more information about DevOps at AWS (What is DevOps).

Some important references

Another AWS service that can used for this purpose is Step Functions. In this link, you'll also find the documentation of the python library that can be executed directly from your Jupyter notebook.

Apache AirFlow is a poweful Open Source tool that can also be integrated with SageMaker. Curious? Just take a look on the SageMaker Operators for AirFlow.

Ah, you have a Kubernetes cluster and want to integrate SageMaker to that and manage the ML Pipeline from the cluster. No problem, take a look on the SageMaker Operators for Kubernetes.

Anyway, there are lots of workflow managers that can be perfectly integrated with SageMaker to do the same job! Pick yours and use your creativity to create your own MLOps plaform!

Pre-Requisites

Services

You should have some basic experience with:

• Train/test a ML model
• Python (scikit-learn)
• Jupyter Notebook
• AWS CodePipeline
• AWS CodeCommit
• AWS CodeBuild
• Amazon ECR
• Amazon SageMaker
• AWS CloudFormation

Some experience working with the AWS console is helpful as well.

AWS Account

In order to complete this workshop you'll need an AWS Account with access to the services above. There are resources required by this workshop that are eligible for the AWS free tier if your account is less than 12 months old. See the AWS Free Tier page for more details.

Scenario

In this workshop you'll implement and experiment a basic MLOps process, supported by an automated infrastructure for training/testing/deploying/integrating ML Models. It is comprised into four parts:

1. You'll start with a WarmUp, for reviewing the basic features of Amazon Sagemaker;
2. Then you will optionally create a Customized Docker Image with your own algorithm. We'll use scikit-learn as our library;
3. After that, you will train the model (using the buil-in XGBoost or the a custom container if you ran the step 2), deploy them into a DEV environment, approve and deploy them into a PRD environment with High Availability and Elasticity;
4. Finally, you'll run a Stress test on your production endpoint to test the elasticity and simulate a situation where the number of requests on your ML model can vary.

Parts 2 and 3 are supported by automated pipelines that reads the assets produced by the ML devoloper and execute/control the whole process.

Architecture

For part 2 the following architecture will support the process. In part 2 you'll create a Docker image that contains your own implementation of a RandomForest classifier, using python 3.7 and scikit-learn. Remember that if you are happy with the built-in XGboost you can skip this part.

1. The ML Developer creates the assets for Docker Image based on Scikit Learn, using Sagemaker, and pushes all the assets to a Git Repo (CodeCommit);
2. CodePipeline listens the push event of CodeCommit, gets the source code and launches CodeBuild;
3. CodeBuild authenticates into ECR, build the Docker Image and pushes it into the ECR repository
4. Done.

For part 3 you'll make use of the following structure for training the model, testing it, deploying it in two different environments: DEV - QA/Development (simple endpoint) and PRD - Production (HA/Elastic endpoint).

Altough there is an ETL part in the Architecture, we'll not use Glue or other ETL tool in this workshop. The idea is just to show you how simple it is to integrate this Architecture with your Data Lake and/or Legacy databases using an ETL process

1. An ETL process or the ML Developer, prepares a new dataset for training the model and copies it into an S3 Bucket;
2. CodePipeline listens to this S3 Bucket, calls a Lambda function for start training a job in Sagemaker;
3. The lambda function sends a training job request to Sagemaker;
4. When the training is finished, CodePipeline gets its status goes to the next stage if there is no error;
5. CodePipeline calls CloudFormation to deploy a model in a Development/QA environment into Sagemaker;
6. After finishing the deployment in DEV/QA, CodePipeline awaits for a manual approval
7. An approver approves or rejects the deployment. If rejected the pipeline stops here; If approved it goes to the next stage;
8. CodePipeline calls CloudFormation to deploy a model into production. This time, the endpoint will count with an AutoScaling policy for HA and Elasticity.
9. Done.

Crisp DM

It is important to mention that the process above was based on an Industry process for Data Mining and Machine Learning called CRISP-DM.

CRISP-DM stands for “Cross Industry Standard Process – Data Mining” and is an excellent skeleton to build a data science project around.

There are 6 phases to CRISP:

• Business understanding: Don’t dive into the data immediately! First take some time to understand: Business objectives, Surrounding context, ML problem category.
• Data understanding: Exploring the data gives us insights about tha paths we should follow.
• Data preparation: Data cleaning, normalization, feature selection, feature engineering, etc.
• Modeling: Select the algorithms, train your model, optimize it as necessary.
• Evaluation: Test your model with different samples, with real data if possible and decide if the model will fit the requirements of your business case.
• Deployment: Deploy into production, integrate it, do A/B tests, integration tests, etc.

Notice the arrows in the diagram though. CRISP frames data science as a cyclical endeavor - more insights leads to better business understanding, which kicks off the process again.

Instructions

First, you need to execute a CloudFormation script to create all the components required for the exercises.

1. Select the below to launch CloudFormation stack.

Region	Launch
US East (N. Virginia)

1. Then open the Jupyter Notebook instance in Sagemaker and start doing the exercises:

   1. Warmup: This is a basic exercise for exploring the Sagemaker features like: training, deploying and optmizing a model. If you already have experience with Sagemaker, you can skip this exercise.
   2. Container Image with a Scikit Classifier: In this exercise we'll create a Docker Image that encapsulates all the code required for training and deploying a RandomForest classifier. If you don't want to create a custom container, skip this section.
      1. Test the models locally: This is part of the exercise #3. You can use this jupyter to test your local WebService, to simulate how Sagemaker will call it when you ask it to create an Endpoint or launch a Batch job for you.
      2. Test the container using a SageMaker Estimator: This optional exercise can be use for understanding how SageMaker Estimators can encapsulate your container and abstract the complexity of the training/tuning/deploying processes.
   3. Train your models: In this exercise you'll use the training pipeline. You'll see how to train or retrain a particular model by just copying a zip file with the required assets to a given S3 bucket.
      1. Check Training progress and test: Here you can monitor the training process, approve the production deployment and test your endpoints.
   4. Stress Test: Here you can execute stress tests to see how well your model is performing.

---

Cleaning

First delete the folowing stacks:

• mlops-deploy-iris-model-dev
• mlops-deploy-iris-model-prd
• mlops-training-iris-model-job

Then delete the stack you created. If you named AIWorkshop, find this stack using the CloudFormation console and delete it.

WARNING: All the assets will be deleted, including the S3 Bucket and the ECR Docker images created during the execution of this workshop.

---

Suggested agenda

• Introduction (1:00)
  • MLStack@AWS, SageMaker concepts/features, MLOps, etc.
• WarmUp - Part 1 (0:30)
• Break (0:15)
• WarmUp - Parts 2,3,4 (0:50)
• Container Image (0:40)
• Lunch (1:00)
• MLOps Pipeline: Train + Deployment (0:30)
• Stress tests + Auto Scaling (0:30)
• Wrap up + discussion (0:20)

Total: 5:45

---

License Summary

This sample code is made available under a modified MIT license. See the LICENSE file.

Thank you!