Observability

Observability is the ability to explain what happened inside a system from emitted evidence. Monitoring asks whether known conditions are healthy; observability lets an engineer diagnose why a specific decision, latency spike, or data anomaly occurred.

The four observability pillars

The core signals are metrics, logs, traces, and events. Each answers a different diagnostic question, and each gains an ML-specific payload:

PillarAnswersML addition
Metricsis the aggregate healthy?score, latency, and fallback-rate distributions
Logswhat happened for this request?prediction metadata and input provenance
Traceswhere did time or failure go?spans across feature → model → policy services
Eventsdid a notable state change occur?model deploys, version promotions, drift alarms

Beyond the generic signals, ML systems need prediction metadata: request ID, model version, feature version, dataset or prompt version, score, threshold, fallback path, latency, and policy decision. In a microservices path, trace context must cross the application, feature service, model-serving service, and downstream workflow so one transaction can be reconstructed end to end.

Artifact: Prediction Event

{
  "event": "prediction_served",
  "request_id": "req-7f31",
  "trace_id": "4bf92f3577b34da6a3ce929d0e0e4736",
  "model_name": "fraud-scorer",
  "model_version": "42",
  "feature_version": "features:v19",
  "score": 0.873,
  "threshold": 0.82,
  "decision": "manual_review",
  "latency_ms": 87,
  "fallback_reason": null
}

This event supports production incident response: responders can filter all decisions made by version 42, find high-latency traces, and compare decisions against later labels. Sensitive raw inputs should not be logged unless the governance and retention controls are explicit.

Failure Modes

High-volume logs are not observability if they cannot answer operational questions. Missing correlation IDs, sampled-away rare failures, and free-text logs make diagnosis slow. Over-logging personal data creates reliability and governance risk even when debugging improves.

References