Observability
Operators should separate liveness from readiness, and track request-rate / error / latency SLIs from Prometheus metrics scraped on the node.
Health endpoints
| Path | Meaning | Typical use |
|---|---|---|
GET /health/live | Process is up | Kubernetes / node liveness — always 200 with { "status": "alive" } when the process responds |
GET /health/ready | Credentials + ledger gateway healthy | Readiness / load balancer — 200 healthy or 503 unhealthy |
GET /health | Lightweight credentials check | Coarse health JSON (credentials on disk) |
GET /health/startup | Startup probe | Credentials present before accepting traffic |
Prefer /health/ready before sending sync or token traffic. Prefer /health/live only to detect a dead process — it does not prove the ledger gateway is reachable.
SLIs
Track these for the node HTTP contract (from /metrics histograms and counters):
| SLI | Definition |
|---|---|
| RPS | Request rate by route group (assets, tokens, health, …) |
| Error % | Share of responses with status ≥ 500 (and optionally 4xx for client-error budgets) |
| p95 latency | 95th percentile request duration for token and asset routes |
Alert on sustained error% or p95 regressions after deploys / on-ledger upgrades — not on single-request noise.
Metrics scrape
GET /metrics exposes Prometheus text for internal scrapers (e.g. Alloy alongside the node process), not partner internet access.