AviralGarg.com

MLOps Design Considerations

Agenda

  1. Introduction and Overview of MLOps
  2. Introduction to AWS and Cloud Computing
  3. Top Products Demo in AWS
    • Model Monitoring
    • ML Solutions Available in AWS

Introduction to MLOps

MLOps Mind Map

MLOps Mind Map

Why Not Run Notebooks in Production?

Meme: "That's the neat part—you don't"

  1. Version Control:
    • Difficult to track changes and perform merges.
    • The cell-based structure complicates versioning.
  2. Execution Order:
    • Out-of-order execution leads to unpredictability.
    • Variables can be defined in any order, causing inconsistent results.
  3. Modular Code:
    • Notebooks often lack modularity, making code reuse challenging.
  4. Testing and Code Reviews:
    • Difficult to implement testing frameworks.
    • Challenging to perform thorough code reviews.
  5. Deployment Complexity:
    • Not designed for production environments.
    • Issues with packaging and deploying notebooks.
  6. Environment Consistency:
    • Dependency management is complex.
    • Environments may be inconsistent between development and production.
  7. Orchestration and Logging:
    • Difficult to orchestrate workflows.
    • Limited logging capabilities for monitoring.
  8. Security Concerns:
    • Notebooks may expose sensitive data.
    • Lack proper authentication and authorization mechanisms.

Importance of MLOps


Netflix Movie Recommender Example

Problem Statement

Scenario 1: Rule-Based System

Architecture Diagram

alt text

Characteristics

DevOps Perspective

Scenario 2: ML-Based System

Architecture Diagram

alt text

Characteristics

MLOps Perspective

Comparing DevOps and MLOps

DevOps

MLOps

Key Differences

Key Takeaways


Deploying Models Without APIs


MLOps System Design

Deployment Strategies

Deploying Models Without APIs

Continuous Improvement in MLOps

Building and Deploying Models

Responsibilities in MLOps

System Design in MLOps

High-Level Overview

alt text

MLOps Tools and Technologies

ETL and Feature Engineering

Standardization in MLOps


MLOps Lifecycle

Phases and Activities

  1. ML Development:
    • Experimentation and prototyping of models.
    • Output: Formalized training procedures.
  2. Training Operationalization:
    • Creating robust training pipelines based on ML engineers’ requirements.
  3. Continuous Training:
    • Regularly updating models with new data.
    • Ensures models remain relevant over time.
  4. Model Deployment:
    • Deploying updated models into production environments.
  5. Prediction Serving:
    • Serving predictions to applications and users.
  6. Continuous Monitoring:
    • Monitoring model performance and detecting issues.
    • Triggers retraining if performance degrades.

Data and Model Management

Differences Between ML Models

Classic ML vs. Deep Learning vs. RAG

Monitoring Model Performance

Variety of MLOps Tools

DevOps and MLOps Integration

DevOps Loop

CI/CD in DevOps

CI/CD/CT in MLOps

Case Study: Uber’s Michelangelo Platform

Amazon SageMaker