Monitor & Drift

Stage: Monitor · What it does: keep an eye on the endpoints you shipped — drift scans, health, usage metrics, and self-healing for deploy-time failures.

Why it matters

Models degrade silently: the world changes, data shifts, and yesterday's model quietly gets worse. Drift monitoring surfaces that before it costs you.
Health + usage give ops the "is it actually working?" answer with real numbers, not vibes.
Self-healing turns deploy failures into retries instead of page-outs.

How to use

Monitor → Drift schedule.
Pick an endpoint and its baseline data path.
Set a scan frequency.
Save — scans run on schedule and flag drift.

Endpoint detail also shows live usage (invocations, latency, errors, CPU/memory) and self-heal status for auto-heal deployments (admin can approve/stop heals).

Real-world examples

🛒 Retail — catching a season shift

A demand-forecast model trained on summer patterns quietly drifts in October. The drift schedule flags the shift in the first week, and the team retrains on autumn data before forecast accuracy visibly tanks.

🏦 Financial services — monitoring a credit model

Regulators want evidence the model's behavior hasn't drifted out of tolerance. A monthly drift scan on the scoring endpoint + data capture gives the compliance team a documented answer and an alerting path.

🧑‍💼 SaaS — ops without firefighting

A support-assistant endpoint's error rate jumps after a bad deploy. Self-heal retries the deploy; if it escalates, the on-call admin gets an actionable banner with the real error and one-click approve/stop.

🏭 Manufacturing — sensor model aging

A quality-prediction model trained on last quarter's sensors starts mismatching. Drift flags the input distribution shift, and the morning pipeline retrains on the freshest telemetry automatically.

Gotchas

A fresh endpoint has no CloudWatch history — the platform bounds idle-burn estimates by creation time so it doesn't misreport "idle" spend.
Drift is only as meaningful as the baseline: choose a baseline period that represents normal for your data.