Introduction

AI Engineering

The model is now a service, and the service has to be built, deployed, and monitored. After this chapter the reader can place feature stores, experiment tracking, deployment, and monitoring on the path from a trained model to a running system, and feature stores continue Data Processing, where raw values were scaled, transformed, imputed, and annotated into the shape a learner expects.

Ops

The pages above are about the method: what it is and how it works. The pages in this half are about how you build, ship, govern, and run the system.

That work splits into five groups. MLOps is the gap between ML in a notebook and ML in production. DataOps starts from asking a warehouse a question in SQL. DevOps opens with the strategy that keeps delivery speed from stalling. DevSecOps puts security inside the delivery loop, not after it. Architecture begins with the signs that a system which looks distributed still behaves like one monolith.

A makeover of these pages was done, and How it was done records the process.

Model platform

The model platform comes first, because everything else in this half serves a model that has to run. Definitions separates the Ops names that get mixed: DataOps, MLOps, GitOps, DevOps, and DevSecOps. MLOps Intro lays out the building blocks and reading list for getting from notebook to production. MLOps Patterns collects the shapes production ML keeps repeating, then two end-to-end systems that use them. MLOps Teams asks who owns that work: machine learning engineers, or a separate MLOps role.

With the owners named, the pages follow the model's path. Feature Stores & Feature Pipelines is the pipeline that builds features and the store that serves the same features to training and serving. ML Experiment Management keeps track of runs that multiply faster than anyone can remember them. Model Formats is how a trained model leaves the notebook in a form another application can read. MLOps Deployment puts that model somewhere that can serve it. MLOps Tools walks the many tools of the stack, from CI and Docker to Kubeflow, workflow runners, and serving. ML Model Monitoring & Alerts keeps a shipped model from going blind. AI As Data treats models as AI tables you query with SQL, and ML Architecture closes the group with end-to-end architecture examples.

Architecture and security

Once the model runs, the system around it has to scale, stay software, and stay secure. System Design is system design interviews, scalability reading, and practice for a service that has to scale. Development Concepts is design patterns, dependency injection, and SOLID, because ML code still has to be software. Problems lists the coupling symptoms of a distributed monolith. On the security side, DevSecOps Definitions gives three definitions of DevSecOps, DevSecOps Tools points at Atlassian's list of the toolchain, and DevSecOps Concepts is the simplified information-security vocabulary to have clear before the tools.

DevOps

The last group is the delivery machinery that every service above depends on. DevOps Strategy points at eight strategies and how to put them in place, and DevOps Tools is the section that holds the tools. Tutorials are two end-to-end walkthroughs: A/B serving, and a training-and-serving stack. Continuous Integration is how code lands continuously. Docker makes a data-science environment run the same way everywhere, and Kubernetes orchestrates those containers. Cloud Objects compares AWS Lambda with Fargate for serverless compute. Key Value DB covers RabbitMQ, ActiveMQ, Kafka, ksqlDB, and ZooKeeper for the messages and state services pass. API Gateway is the single front door for services. Infrastructure As code turns cloud resources into code with Terraform. Logs is the trail production systems leave, ELK is the stack to search and visualize it, and SLO turns uptime promises into measurable objectives with Sloth.