Comparison
ObservabilityvsObserved state
Observability
the incident came in and you had every prompt, output, model version, latency and cost to look at instead of guesses.
Logging, tracing and metrics for LLM features: every call's prompt, output, tokens, latency, cost, model version and outcome. Without it, production incidents reduce to guessing. It is the same discipline you apply to services, with prompts and completions as the payload.
Full entry →Observed state
the spec says three and the status says one, and the gap between those two numbers is the entire incident.
What the system actually reports it has, as opposed to what was asked for. Every orchestrated object carries both, and reading the two side by side is the fastest diagnostic there is: matching means healthy, persistently diverging means a controller is trying and failing and its events will say why. Dashboards that only show desired state look green through an entire outage.
Full entry →