Monitoring

Monitoring helps you answer three questions: is Codex Pooler available, is it keeping up with traffic, and where should you look when something fails?

Start with the admin dashboard for account health, quotas and usage. Add Prometheus and Grafana for trends across requests, application replicas and background jobs.

Set up monitoring

1. Configure metric collection and protect the /metrics endpoint.
2. Import the Grafana dashboard and select your Prometheus data source, namespace and pods.
3. Check a period with known traffic before relying on a graph or alert. An empty panel can mean missing collection or incomplete coverage.

Guide	What it helps you do
Metrics setup	Secure the endpoint, configure a collector and understand which roles emit metrics
Grafana dashboards	Import the starter dashboard and understand its panels and dependencies
Runtime triage	Investigate refusals, connection failures, abandoned reservations and saved resets
PromQL recipes	Query traffic, latency, memory, queues and telemetry relay health
Logs and memory	Correlate logs with requests and investigate memory pressure on any release role

Choose the right evidence

Question	Start here
Can the instance accept traffic?	/healthz, /readyz, scrape health and pod readiness
Are requests slow or failing?	Grafana request, latency and admission panels; then Request logs
Is an account available?	Upstreams, Pools and quota state in the admin UI
Why did a request fail before it appeared in Request logs?	Runtime triage counters and application logs
Did a request use tokens or incur an estimated cost?	Recorded request and accounting data
Is a worker or scheduler running out of memory?	Kubernetes memory/restart metrics and that role’s memory sampler logs

A healthy scrape is not proof that every account has quota. A quiet graph is not proof that no errors occurred. Some failures happen before a request row can be written, and some background events are delivered only on a best-effort basis.

Keep these boundaries in mind

• Pods are not interchangeable. Application memory and admission queues are local to each pod. A cluster total can hide one overloaded replica.
• Client and provider transports can differ. An HTTP/SSE client may use an upstream WebSocket connection. Keep both labels when investigating streams.
• Metrics are not the accounting ledger. Request and token records are the source for recorded usage; charts are a correlation tool.
• Monitor metadata, not content. Do not put prompts, replies, files, audio, images, raw frames or credentials in dashboards, alerts or shared investigation notes.

For account and Pool notifications, use operator alerts. For individual request details, use Request logs.