Building Machine Learning Systems with a Feature Store. Batch, Real-Time, and LLM Systems Jim Dowling

Dostępność produktu

Produkt nie został jeszcze oceniony pod kątem ułatwień dostępu lub nie podano żadnych informacji o ułatwieniach dostępu lub są one niewystarczające. Prawdopodobnie Wydawca/Dostawca jeszcze nie umożliwił dokonania walidacji produktu lub nie przekazał odpowiednich informacji na temat jego dostępności.

Spis treści książki

• Preface
  • Why Did I Write This Book?
  • Target Readers of This Book
  • What This Book Is Not
  • Outline of the Book
  • Conventions Used in This Book
  • Using Code Examples
  • OReilly Online Learning
  • How to Contact Us
  • Acknowledgments
• I. The FTI Pipeline Architecture for Machine Learning Systems
• 1. Building Machine Learning Systems
  • The Anatomy of a Machine Learning System
    • Types of Machine Learning
    • Data Sources
      • Tabular data
      • Event data
      • Graph data
      • Unstructured data
      • API-scraped data
    • Mutable Data
  • A Brief History of Machine Learning Systems
  • MLOps and LLMOps
  • A Unified Architecture for AI Systems: Feature, Training, and Inference Pipelines
    • Classes of AI Systems with a Feature Store
    • ML Frameworks and ML Infrastructure Used in This Book
  • Summary
• 2. Machine Learning Pipelines
  • Building ML Systems with ML Pipelines
    • Minimal Viable Prediction Service
    • Writing Modular Code for ML Pipelines
  • A Taxonomy for Data Transformations in ML Pipelines
    • Feature Types and Model-Dependent Transformations
    • Reusable Features with Model-Independent Transformations
    • Real-Time Features with On-Demand Transformations
    • The ML Transformation Taxonomy and ML Pipelines
  • Feature Pipelines
  • Training Pipelines
  • Inference Pipelines
  • Titanic Survival as an ML System Built with ML Pipelines
  • Summary
• 3. Your Friendly Neighborhood Air Quality Forecasting Service
  • AI System Overview
  • Air Quality Data
  • Exploratory Dataset Analysis
    • Air Quality Data
    • Weather Data
  • Creating and Backfilling Feature Groups
  • Feature Pipeline
  • Training Pipeline
  • Batch Inference Pipeline
  • Running the Pipelines
    • Scheduling the Pipelines as a GitHub Action
    • Building the Dashboard as a GitHub Page
  • Function Calling with LLMs
  • Summary and Exercises
• II. Feature Stores
• 4. Feature Stores
  • A Feature Store for Fraud Prediction
  • Brief History of Feature Stores
  • The Anatomy of a Feature Store
  • When Do You Need a Feature Store?
    • For Context and History in Real-Time ML Systems
    • For Time-Series Data
    • For Improved Collaboration with the FTI Pipeline Architecture
    • For Governance of ML Systems
    • For Discovery and Reuse of AI Assets
    • For Elimination of Offline-Online Feature Skew
    • For Centralizing Your Data for AI in a Single Platform
  • Feature Groups
    • Feature Groups Store Untransformed Feature Data
    • Feature Definitions and Feature Groups
    • Writing to Feature Groups
      • Feature freshness
      • Data validation
  • Data Models for Feature Groups
    • Dimension Modeling with a Credit Card Data Mart
      • Labels are facts, and features are dimensions
      • Feature stores and SCD types
    • Real-Time Credit Card Fraud Detection ML System
      • Data model for our real-time fraud detection ML system
      • Star schema data model
      • Snowflake schema data model
  • Feature Store Data Model for Inference
    • Online Inference
    • Batch Inference
  • Reading Feature Data with a Feature View
    • Point-in-Time Correct Training Data with Feature Views
    • Online Inference with a Feature View
  • Summary and Exercises
• 5. Hopsworks Feature Store
  • Hopsworks Projects
    • Storing Files in a Project
    • Access Control Within Projects
    • Access Control at the Cluster Level Using Projects
  • Feature Groups
    • Versioning
      • Data versioning in feature groups and time travel
      • Versioning feature groups
    • Online Store
      • Time to live
      • Vector index
    • Offline Store (Lakehouse Tables)
      • External feature groups
      • Data statistics
    • Change Data Capture for Feature Groups
  • Feature Views
    • Feature Selection
    • Model-Dependent Transformations
    • Creating Feature Views
    • Training Data as Either DataFrames or Files
      • Random, time-series, and stratified splits
      • Reproducible training data
    • Batch Inference Data
    • Online Inference Data
  • Faster Queries for Feature Data
  • Summary and Exercises
• III. Data Transformations
• 6. Model-Independent Transformations
  • Source Code Organization
  • Feature Pipelines
  • Data Transformations for DataFrames
    • Row SizePreserving Transformations
    • Row and Column SizeReducing Transformations
    • Row/Column SizeIncreasing Transformations
    • Join Transformations
  • DAG of Feature Functions
    • Lazy DataFrames
    • Vectorized Compute, Multicore, and Arrow
    • Data Types
      • Arrays, structs, maps, and tensors
      • Implicit or explicit schemas for feature groups
  • Credit Card Fraud Features
  • Composition of Transformations
  • Summary and Exercises
• 7. Model-Dependent and On-Demand Transformations
  • Feature Transformations
    • Encoding Categorical Variables
    • Distributions of Numerical Variables
    • Transforming Numerical Variables
    • Storing Transformed Feature Data in a Feature Group
  • Model-Specific Transformations
    • Outlier Handling Methods
    • Imputing Missing Values
    • Data Cleaning as Model-Based Transformations
    • Target-/Label-Dependent Transformations
    • Expensive Features Are Computed When Needed
    • Tokenizers and Chat Templates for LLMs
  • Transformations in Scikit-Learn Pipelines
  • Transformations in Feature Views
  • On-Demand Transformations
  • PyTorch Transformations
  • Using pytest
    • Unit Tests
      • Implementing pytest unit tests
      • Running pytest as part of a GitHub Action
    • A Testing Methodology
  • Summary and Exercises
• 8. Batch Feature Pipelines
  • Batch Feature Pipelines
  • Feature Pipeline Data Sources
    • Batch Data Sources
    • Streaming Data Sources
    • Unstructured Data in Object Stores and Filesystems
    • API and SaaS Sources
  • Synthetic Credit Card Data with LLMs
    • A Logical Model for the Data Mart and the LLM
    • LLM Prompts to Generate the Synthetic Data
  • Backfilling and Incremental Updates
    • Polling and CDC for Incremental Data
      • Polling
      • Change data capture
    • Backfill and Incremental Processing in One Program
  • Job Orchestrators
    • Modal
    • Hopsworks Jobs
  • Workflow Orchestrators
    • Airflow
    • Cloud Provider Workflow Orchestrators
  • Data Contracts
  • Data Validation with Great Expectations in Hopsworks
  • Summary and Exercises
• 9. Streaming and Real-Time Features
  • Interactive AI-Enabled Systems Need Real-Time Features
  • Event-Streaming Platforms
  • Shift Left or Shift Right?
    • Shift-Right Architectures
    • Shift-Left Architectures
      • Hybrid streaming-batch architecture
      • Streaming-native architecture
  • Writing Streaming Feature Pipelines
    • Dataflow Programming
    • Stateless and Stateful Data Transformations
    • Apache Flink
    • Feldera
  • Windowed Aggregations
    • Rolling Aggregations
    • Time Window Aggregations
    • Choosing the Best Window Type for Aggregations
    • Rolling Aggregations with Incremental Views
  • Credit Card Fraud Streaming Features
    • ASOF Joins and Composition of Transformations
    • Lagged Features and Feature Pipelines in Feldera
  • Summary and Exercises
• IV. Training Models
• 10. Training Pipelines
  • Unstructured Data and Labels in Feature Groups
    • Self-Supervised and Unsupervised Learning
    • Supervised Learning Requires a Label
  • Root and Label Feature Groups
  • Feature Selection
  • Training Data
    • Splitting Training Data
    • Reproducible Training Data
  • Model Training
    • Model Architecture
    • Checkpoints to Recover from Failures
    • Hyperparameter Tuning with Ray Tune
    • Distributed Training with Ray
    • Parameter-Efficient Fine-Tuning of LLMs
    • Credit Card Fraud Model with XGBoost
    • Identifying Bottlenecks in Distributed Training
  • Model Evaluation and Model Validation
    • Model Performance for Classification and Regression
    • Model Interpretability
    • Model Bias Tests
    • Model File Formats and the Model Registry
    • Model Cards
  • Summary and Exercises
• V. Inference and Agents
• 11. Inference Pipelines
  • Batch Inference Pipelines
    • Batch Predictions for a Time Range
    • Batch Predictions for Entities
    • Scaling Batch Inference with PySpark
    • Data Modeling for Batch Inference
    • Batch Inference for Neural Networks
  • Batch Inference for LLMs
  • Online Inference Pipelines
    • Ensure Offline-Online Consistency for Libraries
    • Model Deployments with FastAPI
    • LLM Deployments
    • Deployment API for Models and Feature Views
  • Model-Serving Frameworks with KServe
  • Performance and Failure Handling
    • Mixed-Mode UDFs
    • Native UDFs and Log-and-Wait
    • Handling Failures in Online Inference Pipelines
    • Model Deployment SLOs
  • Inference with Embedded Models
    • Embedded AI-Enabled Applications
    • Stream-Processing AI-Enabled Applications
    • UIs for AI-Enabled Applications in Python
  • Summary and Exercises
• 12. Agents and LLM Workflows
  • From LLMs to Agents
    • Prompt Management
    • Prompt Engineering
    • Context Window
    • Agents and Workflows with LlamaIndex
  • Retrieval-Augmented Generation
    • Retrieval with a Document Store
    • Retrieval with a Feature Store
    • Retrieval with a Graph Database
  • Tools and Function-Calling LLMs
  • Model Context Protocol
  • Agent-to-Agent (A2A) Protocol
  • From LLM Workflows to Agents
    • Planning
    • Security Challenges
    • Domain-Specific (Intermediate) Representations
  • A Development Process for Agents
  • Agent Deployments in Hopsworks
  • Summary and Exercises
• VI. MLOps and LLMOps
• 13. Testing AI Systems
  • Offline Testing
  • From Dev to Prod
  • Automatic Containerization and Jobs
    • Environments and Jobs in Hopsworks
    • Modal Jobs
  • CI/CD Tests for AI Systems
    • Feature Pipeline Tests
    • Training Pipeline Tests for Model Performance and Bias
    • Testing Model Deployments
    • A/B Tests for Batch Inference
    • Evals for Agents
      • LLM-assisted synthetic eval generation
      • Historical evals require point-in-time correct RAG data
  • Governance
    • Schematized Tags
    • Lineage
    • Versioning
    • Audit Logs
  • Summary and Exercises
• 14. Observability and Monitoring AI Systems
  • Logging and Metrics for ML Models
    • Logging for Batch and Online Models
    • Metrics for Online Models
    • Metrics for Batch Models
  • Monitoring Features and Models
    • Data Ingestion Drift
    • Univariate Feature Drift
    • Multivariate Feature Drift
    • Monitoring Vector Embeddings
    • Model Monitoring with NannyML
    • When to Retrain or Redesign a Model
  • Logging and Metrics for Agents
    • From Logs to Traces with Agents
    • Error Analysis
      • Log viewer and feedback
      • Curating evals
    • Guardrails
    • Online A/B Testing
    • Jailbreaking and Prompt Injection
    • LLM Metrics
  • Summary and Exercises
• 15. TikToks Personalized Recommender: The Worlds Most Valuable AI System
  • Introduction to Recommenders
  • A TikTok Recommender with the Retrieval-and-Ranking Architecture
  • Real-Time Personalized Recommender
    • Feature Pipelines
    • Training Pipelines
      • Two-tower embedding model
      • Building the vector index of videos
      • Ranking model
    • Online Inference Pipeline
  • Agentic Search for Videos
  • The Dirty Dozen of Fallacies of MLOps
  • The Ethical Responsibilities of AI Builders
  • Summary
• Index

pokaż cały spis treści

ukryj recenzje