"using sagemaker to build and deploy ml models in production" - MJ Berends AWS Chicago user group June 6 2019
0 likes409 views
The document outlines the use of machine learning (ML) in production at ActiveCampaign, detailing the process from defining tasks to deploying models using Amazon SageMaker. It emphasizes the importance of data processing, model training, and evaluation while providing a structured overview of the ML workflow and the tools used, such as Jupyter notebooks and MLflow. Additionally, it highlights the collaboration across teams to successfully implement ML solutions.
"using sagemaker to build and deploy ml models in production" - MJ Berends AWS Chicago user group June 6 2019
- 2. MJ Berends Staff Engineer @ ActiveCampaign Data and analytics architect with a love for building large-scale systems that distill the essence from highly complex and idiosyncratic data.
- 3. ActiveCampaign helps growing businesses meaningfully connect and engage with their customers. We go beyond marketing automation to enable businesses to optimize their customers’ experiences. 71,000+ Customers 400+ Employees 150+ Countries
- 4. Roadmap 1. What even is Machine Learning? 2. A view of the ML process 3. A quick tour of SageMaker 4. A simple model with SageMaker 5. The production ecosystem
- 5. What is SageMaker? SageMaker helps data scientists and developers to build, train, and deploy machine learning models into a production-ready hosted environment.
- 6. What even is machine learning? A machine learns a task by reducing its performance error
- 7. What even is machine learning? Define a task 1 Define Performance Error 2 Define Starting Place, Incremental Improvement, Stopping Place 3 TASK REWARD PATH
- 9. Find a Peanut Butter Sandwich 1. Think of a canonical example 2. Enumerate rules that describe that example a. Has bread b. Has peanut butter c. Is edible 3. Use those features to make guesses 4. Update your rules as you find counter-examples Rules
- 10. Find a Peanut Butter Sandwich Statistics 1. Find sandwich examples and label them as “good” or “bad” 2. Identify features that make it likely that something is a sandwich 3. Use an algorithm to make guesses 4. Gradually change the features until your guesses improve
- 11. The Machine Learning Process Query the data warehouse to uncover preliminary patterns Data Sleuthing Select a few promising approaches to modeling Model Brainstorming Train the model candidates and inspect the performance metrics Model evaluation Clean and transform the data for model training Training Data Creation Pick the best-performing model Model Selection Onward to production! App Data Warehouse Experiment management app
- 12. The Implementation at 20,000 Feet Model ServingData Ingestion Application Data Warehouse Training Data Model training MySQL Snowflake S3 SageMaker Python MLFlow C + Model Artifacts Webserver C + Training Instances Serving Instances SageMaker Python API Gateway S3
- 13. Model Exploration and Training Data Scientist Exploration Jupyter Notebook Jupyter Notebook SageMaker Data Lake Table 1 Table 2 MLFLow Artifact Store File Product DBs API SQL Query Airflow Sagemaker Training Env Task S3 Container
- 14. Model Exploration and Training Data Scientist Exploration Jupyter Notebook Jupyter Notebook SageMaker Data Lake Table 1 Table 2 MLFLow Artifact Store File Product DBs All data used in exploration comes from Data Lake The product data is exported to Warehouse periodically. API SQL Query Airflow Sagemaker Sagemaker manages container tasks Training Env Task S3 Container
- 15. Model Exploration and Training Data Scientist Exploration Jupyter Notebook Jupyter Notebook SageMaker Data Lake Table 1 Table 2 Training Env TaskMLFLow Artifact Store S3 File Product DBs Models are built in Jupyter notebooks. MLFlow manages deployment and storage API SQL Query Airflow Sagemaker
- 16. Model Exploration and Training Data Scientist Exploration Jupyter Notebook Jupyter Notebook SageMaker Data Lake Table 1 Table 2 MLFLow Artifact Store File Product DBs MLFlow manages deployment and storage API SQL Query Airflow Sagemaker Training Env Task S3 Container
- 17. Model Exploration and Training Data Scientist Exploration Jupyter Notebook Jupyter Notebook SageMaker Data Lake Table 1 Table 2 MLFLow Artifact Store File Product DBs API SQL Query Airflow Sagemaker Training Env Task S3 Container Sagemaker manages container tasks
- 18. App 1 App 2 App 3 Endpoint Model Client Applications Model Serving Users Production Account ECR S3 Sage- maker Data Science Account Endpoint Model Endpoint Model Model Serving
- 19. So, how do you use SageMaker? Publishing a new model to Sagemaker is a three-step process, consisting of: 1. Creating a Model 2. Creating an Endpoint Configuration 3. Creating an Endpoint
- 20. What’s in a Model? Network VPC Access Controls Execution Role Container Image S3 ECR VPC EndpointSecurity Groups Subnets Model Data URL
- 21. What’s in an Endpoint Config? Instance Count Instance Type Model Variant Weight How many instances should but spun up at the start? How big should the instances be? How should traffic be routed among models?
- 22. What’s in an Endpoint? Endpoint Config Instances All the details for spinning up a fleet of model versions One or more instances that run model containers Autoscaling via Application Autoscaling Also... Monitoring via CloudWatch
- 23. Lessons Build relationships across teams Integrate with the product early Focus on the modeling ecosystem Expect requirements to change Iterate Always
- 24. THANK YOU! Q&A
- 27. Model Exploration Data Scientist Exploration Jupyter Notebook Jupyter Notebook SageMaker Data Lake Table 1 Table 2 MLFLow Artifact Store File Product DBs Models are built in Jupyter notebooks. MLFlow manages deployment and storage All data used in exploration comes from Data Lake The product data is exported to Warehouse periodically. API SQL Query Airflow Sagemaker Sagemaker manages container tasks Training Env Task S3 Container