The ML Lifecycle management process is quickly becoming the bottleneck for a lot of ML projects. With MLflow’s newest release, and its enhanced integration with Azure Machine Learning, this process is now showing the right promise and capabilities on Azure. In this talk, we intend to take a tour of the integration details and how MLOps is now becoming a strength of the platform. We’ll talk about versioning, maintaining run history, production pipeline automation, deployment to cloud and edge, and CI/CD pipelines with MLOps as the backdrop.
Be prepared for an interactive conversation as we intend to seek a lot of feedback on the integration and capabilities being lit up.
4. Azure Databricks
Fast, easy, and collaborative Apache Spark™-based analytics platform
Built with your needs in mind
Role-based access controls
Effortless autoscaling
Live collaboration
Enterprise-grade SLAs
Best-in-class notebooks
Simple job scheduling
Seamlessly integrated with the Azure Portfolio
Increase productivity
Build on a secure, trusted cloud
Scale without limits
5. Azure Machine Learning service
Bring AI to everyone with an end-to-end, scalable, trusted platform
Built with your needs in mind
Support for open source frameworks
Managed compute
DevOps for machine learning
Simple deployment
Tool agnostic Python SDK
Automated machine learning
Seamlessly integrated with the Azure Portfolio
Boost your data science productivity
Increase your rate of experimentation
Deploy and manage your models anywhere
6. Large retail customer: Use case + Persona
Customer Ask: We need to build a unified platform to support a large globally diverse team of Data Engineers,
Data Scientists and AI Developers for their big data, deep learning projects.
These projects will help us predict and reduce churn, increase retention, and grow revenue
Data Scientists want to use
open source frameworks like
PyTorch & TensorFlow with
GPU and CPU for training.
They are familiar with MLflow
for managing ML lifecycle.
ML Engineers want to
integrate the ML models in to
applications via a scalable web
service.
They do not want to manage
the infrastructure.
Recommendation (preferred by the customer):
Use Azure Databricks for data prep
Use Azure ML with MLflow on Azure Databricks for training OR
Use Azure ML with MLflow in Notebook VM with remote Azure ML compute for training
Use Azure ML for Model management and MLOps
Data Engineers prefer to stay
in Spark for distributed data
processing on PB scale data.
They do not want to manage
the infra for data preparation
7. Experiments
Local machine
Virtual machine
Azure ML Compute
Azure Databricks
Experiments and Metrics Logging
Azure Machine
Learning Workspace
Experiments and
Metrics Tracking
Metric
s
Artifacts
Logging API
Tracking URI
Demo summary – MLflow with Azure ML Experimentation
9. How to get started
Install
• PyPi package: azureml-
mlflow
Set
• Set Azure ML workspace
• Set MLflow tracking URI to
Azure ML
Go
• Run your MLflow experiment
• Track your results in Azure ML
• Deploy trained model to Azure