19. MLOps

ML process

Main Challenges of ML

• publishing an ML model is not enough
• managing and published ML models is as important as developing them

 

ML Process

 

Phase 1: Research / Experiment

• Can we use ML to solve this?
• scientific projects / Proof-of-concepts (PoC)
• focus on model performance

Phase 2: Operational

• How do we implement this method at scale?
• After PoC, bringing your ML models to production
• Migration of existing models into ML platform

---

MLOps

MLOps = ML + Dev + Ops

• Collaborataive and experimental in nature
• Automate as much as possible
• Continuous improvement of ML models
• Standardize and scale
• 👉 allow us to focus on business logic

 

Components of MLOps

Serving: deploy model to production in real world

• In Research: they don't serve model
• Batch serving: batch process, higher throughput, higher latency
• Online serving: lower throughput, lower latency

Management: monitoring and recording the training results and artifact

• artifact: outcomes from training (ex. model file)
• In production: can get user's feedback

Feature store: store ML features that are commonly shared

 

Data Validation

• Data drift, model drift, concept drift
• data & model in research != data & model in production
• static model v.s. refreshed models
  • refreshed model: you can maintain your model quality

 

Continuous Training

• do continous training through automation

Others

• Model analysis, AutoML, Infrastructure management, Monitoring ...

 

Why MLOps

Agility

• continuous and faster deliveries
• fastser modifications and bug-fixing

Experiments

• faster and controlled experiments
• faster integration of successful experiments to other environments

Scalability

• ease integration of new ML model
• standardization of code
• lower operational costs

Time to market

• reduced time-to-market
• faster planning and delivery expectations

buisness owners

• strong collaboration
• improve iterations

 

MLOps practices

Difference btw DevOps and MLOps

• MLOps: end-to-end ML lifecycle management
  • including data provenance, datasets, models, hyperparameters, metrics, workflows

MLOps Practices

• Training pipeline
• Deployment pipeline
• Those two pipeline goes parallel in practice

 

Amazon SageMaker as an MLOps tool

Amazon SageMaker

• most complete, end-to-end ML service
• Integrated workbench, Managed infrastructure, Managed tooling
• Automate ML workflows to scale model development
• Build CI/CD pipelines for ML to accelerate model deployment
• Catalog informations for traceability and reusability
• Track lineage for troubleshooting and compliance
• Maintain accuracy of predictions after models are deployed
• Enhance governance and security