BHANUJBHANUJ
v1.2.2

Observability

AIGP provides metrics, logs, and traces so operators can see how jobs, tests, and decisions behave in production.

The 3 Signals

Every subsystem is instrumented for metrics, logs and traces. All three share correlation identifiers so a single request can be followed across control planes.

Metrics

  • Job throughput, latency and retry counts.
  • Evaluation duration and cache hit rate.
  • Reconciliation lag between relational and graph state.

Structured Logs

Logs are emitted as structured JSON with a stable schema and correlation information, making operational behavior queryable rather than free text. Keep protected content and high-cardinality tenant data out of logs.

json
{
  "level": "info",
  "event": "decision.committed",
  "decision_id": "decision_991",
  "trace_id": "b7ad...e2",
  "duration_ms": 42
}

Traces

Distributed traces span the full lifecycle — intake, evaluation, decision and projection — so latency can be attributed to a specific stage rather than guessed at.

AI Governance Platform / governed event stream
trace / b7ad…e2event order
  1. 01
    record intake

    A correlation ID begins at the governed request boundary.

  2. 02
    attach evaluation span

    Provider latency and metric work remain linked to the same trace.

  3. 03
    commit decision event

    The decision carries its policy, evidence and duration context.

  4. 04
    export telemetry

    Metrics, structured logs and trace spans reach the existing backend.

governed outcome

TRACE LINKED

One correlation ID · full lifecycle visible · latency attributable

OpenTelemetry

OpenTelemetry is the deployment-owned path for operational traces and metrics. For PostgreSQL client work, create low-cardinality client spans and prefer safe operation summaries over raw SQL or query parameters.

Safe Observability Data

Do not place prompts, responses, tool payloads or results, credentials, tokens, tenant or project identifiers, evidence references, raw SQL parameters, or private agent content in logs, trace attributes, or metric labels. Those signals help run the platform; they are not another evidence store.

Product telemetry is a separate, optional, privacy-preserving feature. Read Telemetry & Privacy to see its defaults, allow-listed data, preview endpoint, and disable switch.