Observability

Atlas observability connects runtime events to an operator decision. Metrics quantify scope, traces localize request paths, and structured logs explain discrete events and policy outcomes. None is sufficient alone. Release, dataset, profile, route class, and time must remain joinable across the evidence window.

Correlation spine

flowchart TD
    Event[Request or lifecycle event] --> Request[Request + trace identity]
    Event --> Release[Runtime + dataset identity]
    Event --> Context[Profile + instance + route class]
    Request --> Logs[Structured logs]
    Request --> Traces[Trace and spans]
    Release --> Metrics[Metrics and SLO window]
    Context --> Metrics
    Logs --> Window[Correlated evidence window]
    Traces --> Window
    Metrics --> Window
    Window --> Decision{admit, drain, mitigate, promote}

High-cardinality request and trace identifiers belong in logs and traces. Metrics use bounded dimensions defined by their contracts. A signal that cannot be joined to the release and observation window is diagnostic material, not promotion evidence.

Qualify each signal path

flowchart LR
    Instrument[Instrumented] --> Emit[Emitted]
    Emit --> Deliver[Delivered]
    Deliver --> Query[Queried]
    Query --> Exercise[Rule + notification exercised]
    Exercise --> Retain[Decision evidence retained]

Level	Evidence	Claim supported
declared	Registry, rule, dashboard, or drill validates	Intended signal, dimensions, and owner are specified
emitted	A known event produces the expected signal	Instrumentation is active for that path
delivered	A backend query returns the signal with expected identity and time	Transport, ingestion, and retention worked for the window
exercised	A controlled condition reaches rule, notification, view, and correlated context	The response path worked in the named environment
retained	Immutable evidence binds the exercise to release, dataset, profile, and time	The decision can be reconstructed after live data expires

Qualification is per required path. A latency metric cannot establish audit completeness; one delivered trace cannot establish a population error rate. Promotion policy must name the minimum level for every required signal.

Signal-loss decisions

Missing boundary	Surviving evidence can establish	Claim that remains blocked
metrics delivery	Individual failures from logs, traces, probes, and clients	Population rate, saturation trend, or SLO compliance
trace export	Rates and event classes from metrics and logs	Causal localization across runtime and dependencies
centralized logs	Aggregate health and sampled paths	Complete event, audit, or policy-decision history
alert notification	Rule state and direct backend queries	That the assigned owner received and acted on the page
dashboard rendering	Raw queries and validated panel definitions	That the operator view rendered the same window
all telemetry backends	Clients, Kubernetes, pod output, and integrity checks	Promotion, security closure, or absence of hidden failures

Restoring a path is not retrospective evidence for its blind interval. Record source and collection clocks, the gap, emergency observation methods, and the time normal evidence resumed. Start a new continuous qualification window when policy requires one.

Contracted surface

The checked-in contracts define 15 HTTP endpoint entries, 39 required metrics, structured event fields, six request-path spans, ten lifecycle span identities, 20 governed alerts, and two telemetry-path drills. The generated telemetry index proves that these contracts are discoverable. It does not prove that a deployed collector, backend, rule engine, dashboard, or notification route is working.

Route by operating question

Question	Read
May an instance receive traffic?	Health, Readiness, and Drain
Which runtime path is failing?	Logging, Metrics, and Tracing
What must a structured event retain?	Logging Contracts
Which metric dimensions are governed?	Metrics Packages
Is the service within budget?	Service Objectives and Error Budgets
How does context cross process boundaries?	Tracing Pipelines
Which signal requires action?	Alert Rules
Which view explains impact?	Dashboards and Panels
Can the evidence path survive failure?	Telemetry Drills
What must accompany a decision?	Operational Evidence Reports

Preserve the observation window, release and dataset identities, rule and dashboard revisions, representative trace IDs, structured events, gaps, and decision record. Missing telemetry is an operating finding, not an empty pass.