Learn: Observability

How the platform sees itself — how a slow request, a crash loop, or a runaway LLM bill becomes findable, and how little a team has to do to make it so. A whole-picture tour, then focused deep dives.

Written for platform engineers and any developer who’s debugged from a dashboard. Most of it is platform-injected, so a developer gets the payoff without the plumbing. It helps to have read Foundations (the cluster this runs on) and Delivery, and to know roughly what a metric, a log, and a trace are.

Read in this order

1. Orientation — the one idea, “the platform observes your workload for you,” and a tour: the four signals → zero-code collection (Beyla eBPF) → the self-hosted LGTM+P stack → correlation (the jump) → acting on signals (SLOs, pages, cost) → observing the agents. Includes a straight status of what’s live and what’s deferred: per-team isolation is enforced hub-first, and traces/profiles isolation is deferred.
2. Reference — the dense lookup: versions, the stack, tenancy, the collectors, SLOs/alerts/ cost, agent-obs, the status ledger, gotchas.

Go deep

• The stack & storage — the LGTM+P backends, S3, Grafana + SSO, the X-Scope-OrgID tenancy model (and why isolation is really the network), hub-and-spoke, and why self-hosted.
• Collection & instrumentation — every collector, the eBPF zero-code story (Beyla), the instrumentation ladder (Layer 0/1/2), and platform-injection.
• Correlation & the team experience — the metrics↔traces↔logs↔profiles jumps, the per-team overview dashboards, and the honest per-team-isolation story.
• SLOs, alerting & cost — burn-rate SLOs, curated alerts + triage-agent owner-routing, and cost-as-a-signal (OpenCost + true-cost).
• Agent observability — the OTel GenAI semantic conventions, the agent span tree, and the human-verdict eval loop.

Then

• The substrate it runs on: Foundations. The alert-routing agent: Identity & access (owner resolution). Author obs: the observability-authoring skill.