Observability & Platform Health
Master observability for probabilistic systems — the three pillars, user-facing observability, platform-level monitoring with SLOs, cost observability and showback, and incident response patterns unique to AI.
Observability & Platform Health
The telemetry that makes AI platforms operable
Master observability for probabilistic systems — the three pillars, user-facing observability, platform-level monitoring with SLOs, cost observability and showback, and incident response patterns unique to AI.
Modules in This Level
6.1 Observability for AI-Powered Platforms
Understand why infrastructure observability is insufficient for an AI platform — a healthy server can host a degraded model — and learn the three-layer stack: infrastructure, application services, and AI-specific signals like agent completion rate and first-turn success rate.
6.2 Logs, Metrics, Traces — The Three Pillars
Learn how logs, metrics, and traces serve distinct roles — event audit trail, aggregated rate and latency signals, and per-request cross-service flow — and the sampling and retention policies that keep per-session log volume from becoming an unmanageable storage cost at scale.
6.3 User-Facing Observability
Design layered error presentation for Make's two audiences — raw stack traces for developers, AI-translated plain-language for builders — and the confidence gating and fix-confirmation patterns that prevent the LLM interpreter from modifying code on a wrong diagnosis.
6.4 Platform-Level Monitoring
Build Make's platform health dashboard using five golden signals — preview success rate, mean time to preview, agent session completion, build pass rate, deployment success — and how error budgets turn reliability tradeoffs into math rather than organizational politics.
6.5 Cost Observability and Showback
Attribute Make's infrastructure costs across compute, model inference, storage, and network using resource tagging and per-session AI cost logging — so you can answer 'where is the $2M going' with feature-level precision and make model tier tradeoffs with unit economics.
6.6 Incident Response for AI Systems
Learn the scope-severity-surface triage framework for AI incidents — determining whether a build success rate drop is model regression, registry degradation, runtime drift, or a memory leak — and how synthetic monitoring canaries catch degradation before support tickets arrive.
6.7 Level 6 Checkpoint
Apply Level 6 observability knowledge to a CTO-level scenario: design the cost attribution and monitoring system that answers "is our $2M/month AI infrastructure spend efficient or wasteful" with feature-level precision, error budget reasoning, and actionable unit economics.