Observability and reliability
Instrument a service with OpenTelemetry-spec traces, metrics, and logs, understand why agent traces are kept as evidence, and reason about reliability through SLIs, SLOs, error budgets, and the policy that turns a spent budget into a release freeze.
Units5
Duration26 min
Levelintermediate
By the end of this module, you'll be able to:
- Describe an AI-native telemetry model - OpenTelemetry-spec traces, metrics, and logs under one correlation, bound to a canonical model and queryable from MCP.
- Explain what AI/agent observability adds - the agent trace, token and cost, eval drift, and inference records kept as responsible-AI evidence.
- Define an SLI, an SLO, and an error budget, and compute a budget from an SLO and a window.
- State the error-budget policy - the rule that converts a spent budget into a per-service feature freeze - and why the SLA is always looser than the SLO.
- Read the observability and reliability standards as the source of truth and follow them for the detail.
Prerequisites
- Familiarity with running a web service and reading its logs.
- The "Apply the AGENTS floor" module, or equivalent estate familiarity, is recommended.
Units
- 01Introduction4 min
- 02Instrument for observability7 min
- 03Reliability and SLOs8 min
- 04Knowledge check4 min
- 05Summary3 min
Related
- Design and version an API Shape a clean, consistent API against a clear API design standard - resources, standard methods, typed errors, cursor pagination, idempotency - then version and deprecate it under a versioning policy so no consumer is ever broken without consent.
- Apply the naming standards Name a new identifier, file, repo, or branch the way an AI-native org does - keyword-first, no redundant prefix, cased by role - and tell a functional keyword apart from a product codename, so your first artifact passes the naming gate by default.
- Read the pipeline that ships your change Follow a change from a pull request through the canonical CI stage set to a gated production deploy - the ordered gates a pipeline runs, why CI is the real gate, and how the deployment model promotes one signed artifact dev to staging to prod.