Products
Gateway Observability
Gateway records telemetry from the same path that governs the request. That makes gateway telemetry more useful than backend-only metrics because it includes identity, organization, endpoint, routing, policy, latency, and error context.
What Gateway Emits
Gateway deployments can emit:
- request counts and error counts,
- endpoint latency and backend latency,
- command or query timing where protocol support exists,
- policy allow, warn, block, and redact decisions,
- authentication and RBAC outcomes,
- connection and pool pressure,
- migration progress and divergence,
- model request and token usage where configured,
- traces and structured logs.
Why Gateway Telemetry Matters
Backend telemetry tells you what happened inside a database or service. Gateway telemetry tells you who caused it, which endpoint route was used, what policy did, and how much time was spent in the gateway versus the backend.
This is especially useful for:
- identifying risky commands before they become incidents,
- finding slow or expensive AI/tool requests,
- separating gateway latency from backend latency,
- proving RBAC and policy behavior during audits,
- monitoring migration phases,
- and detecting traffic patterns such as hot keys, N+1 access, missing TTLs, and high fanout.
Operational Dashboards
At minimum, production dashboards should show:
| Dashboard area | Signals |
|---|---|
| Traffic | Request rate, throughput, payload size, connection count. |
| Reliability | Error rate, rejected requests, backend failures, policy blocks. |
| Latency | p50, p95, p99, p99.9, gateway overhead, backend time. |
| Capacity | CPU, in-flight requests, queue depth, lane or pool waiters. |
| Security | Auth failures, RBAC denials, sensitive data findings, dangerous command attempts. |
| Product workflows | Migration phase, AI Orchestration tool calls, workflow runs, approval state. |
Request Analysis
Request analysis can capture richer context than hot-path counters alone. Use it to understand:
- repeated query patterns,
- hot keys and high fanout,
- dangerous command attempts,
- large responses,
- PII or secret exposure,
- and recommendations tied to observed traffic.
Do not turn every deep analysis signal into an always-on hot-path requirement. High-volume deployments should keep hot-path telemetry cheap and move deeper inspection into sampled or bounded paths.