Site Reliability Engineering (SRE)

The discipline of operating production systems for reliability — treating operations as a software problem, governed by SLOs (service-level objectives) and practised through incident management, postmortems, runbooks, and anomaly detection. A pillar of platform-ops.

In the sources

google-sre-agentic-ai shows Google SRE folding aiops into the full SRE workflow — runbook generation, TimesFM-based anomaly detection, comms consolidation, postmortem drafting, and incident investigation — while insisting on SLOs + fallbacks, identity/permissions, explainability, and auditability. The mandate there is telling: even AI-augmented SRE keeps the human accountable and the system constrained.

The canonical practice (Google SRE)

The discipline’s first-party definition is the google-sre-book: SRE began at Google in 2003 under Ben Treynor Sloss and is “what happens when you ask a software engineer to design an operations team.” Three load-bearing mechanisms sit under the practice:

• Error budgets — the reliability target sets a permitted-unavailability budget (a 99.99% SLO ⇒ 0.01% downtime); launches spend it, and exhausting it freezes risky changes. This is the formal lever resolving the dev-velocity-vs-stability tension that service-level-objectives quantifies.
• The 50% toil cap — operational work (tickets, on-call, manual tasks) is capped at half an SRE’s time so the rest funds engineering; this deliberately decouples headcount from service growth.
• Change discipline — ~70% of outages trace to changes, so progressive rollouts + fast detection + safe rollback are core, and on-call targets ≤~2 events per shift to avoid pager fatigue.

These predate the AI augmentation above: aiops accelerates the workflow, but the error-budget / toil / change-management backbone is what it operates within.

Relation to neighbours

• Investigates over observability signals and the service-topology (netflix-service-topology) to locate a failure’s blast radius.
• Overlaps but does not equal platform-engineering: SRE owns reliability of what runs; platform engineering owns the platform it runs on (kubernetes-integration-tax).

Related

google-sre-book · service-level-objectives · dora-metrics · observability · service-topology · aiops · platform-ops · google-sre-agentic-ai