Google revises Site Reliability Engineering book for AI era

Google has released a rewritten second edition of its Site Reliability Engineering book, marking ten years since the first version was published.

The revised edition reflects changes in software operations since the original appeared, including the spread of AI, microservices, multi-cloud deployments, serverless computing and platform engineering.

The first edition set out Google's internal approach to managing large-scale production systems and became widely used by engineering teams beyond the company. The new version revisits those ideas in the context of software environments that change more frequently and depend on increasingly complex systems.

When the first book was published, many organisations were focused on early cloud migrations and the shift away from monolithic systems. Google argues that the central challenge now is maintaining reliability as software is delivered faster and across larger, more fragmented infrastructures.

Operational shifts

The second edition describes how core SRE practices have changed over the past decade. In service level objectives and error budgets, teams have moved from basic availability measures and manual policy decisions towards user-focused targets, dynamic burn-rate alerting and automated enforcement, according to Google.

It also points to changes in monitoring and observability. Earlier approaches often relied on static dashboards, basic logging and manual alert triage, while newer methods use distributed tracing, telemetry standards and automated correlation to identify likely root causes more quickly.

Automation is another area highlighted in the book. Operational work has shifted from custom scripts for individual tasks to declarative infrastructure tools, internal developer platforms and systems designed to recover from faults with less manual intervention, Google says.

The revised edition also covers reliability testing, where scheduled game days and synthetic load tests have increasingly been supplemented by continuous resilience checks, automated pipeline testing and production canary verification. In capacity planning, it says static forecasting and manual provisioning have given way to auto-scaling, closer links with cloud cost management and specific optimisation work for GPU and TPU workloads.

AI focus

A major addition is a section on the relationship between SRE and artificial intelligence. Google frames that relationship in two ways: the work needed to keep AI infrastructure reliable, and the use of AI tools by reliability teams in day-to-day operations.

According to Google, operating large language models and other foundation models creates new reliability demands. These include managing large GPU and TPU clusters, handling network topology issues, monitoring non-deterministic outputs and variable latency, and maintaining stable pipelines for model training, fine-tuning and data processing.

The book also examines how AI tools can be applied within operations teams. Generative AI can assist with incident triage and postmortem creation, while machine learning models can be used to detect resource bottlenecks before they lead to service disruption, Google says. Engineers are also using natural language queries to interrogate system telemetry.

Culture remains

Alongside technical changes, the second edition returns to one of the original text's central ideas: reliability depends on organisational culture as well as engineering methods. Google says the new edition places renewed emphasis on the human side of operations and on building teams and processes around reliability goals.

It also says organisational models for SRE have broadened over the past decade. Instead of relying only on dedicated teams attached to specific applications, some businesses now use hybrid structures, integrate SRE work with platform engineering and embed reliability specialists within product teams.

The updated book is aimed at readers ranging from newcomers to experienced practitioners, Engineering Managers and Technology Executives. Its publication underlines how the discipline has evolved from a specialised operating model associated with large internet platforms into a broader framework for managing increasingly complex production systems.

Modern production environments operate under constant change, with deeper dependencies and more severe failure modes than in the past, Google says.