Observability
Atlas exposes a Prometheus /metrics endpoint and structured logs, plus a terminal inspection tool
for watching a fleet live — the stand-in for the web console (which is its own later milestone).
Metrics
Section titled “Metrics”The gateway serves Prometheus /metrics, including a queue-depth / utilization signal usable for
external autoscaling (e.g. an ASG scaling policy). The endpoint requires an admin-scoped key, so
configure your scraper to send it — mint one with atlas keys create --admin and pass it as a
Authorization: Bearer <key> (or x-api-key) header in the Prometheus scrape config. Then build
dashboards as usual.
CLI inspection
Section titled “CLI inspection”atlas status # a snapshot of the fleetatlas top # a live view — run over SSH on the gateway boxatlas top is the “see what the fleet is doing” view from the terminal: workers, deployed models,
in-flight requests, and load.
Backpressure signals
Section titled “Backpressure signals”Under load Atlas sheds with retryable codes rather than timing out — 429 rate_limit_error
(momentarily full) and 529 overloaded_error (no live replica right now), both carrying
Retry-After. Watching these alongside queue depth tells you when to add capacity. See
API compatibility and
Scaling & operations.