Admission control metrics

The admission-control metrics show why requests are rejected and how close each model is to its configured occupancy and concurrency limits.

Metrics

Metric	Type	Labels	Meaning
`hivenet_router_admission_rejections_total`	Counter	`reason`, `model`	Requests rejected by an admission gate
`hivenet_router_admission_occupancy_tokens`	Gauge	`model`	Current weighted in-flight token sum
`hivenet_router_admission_budget_tokens`	Gauge	`model`	Effective global budget: `HIVENET_ROUTER_ADMIT_FRACTION × admit_budget_tokens × healthy_replicas`
`hivenet_router_admission_inflight_requests`	Gauge	`model`	Current in-flight request count
`hivenet_router_admission_max_inflight`	Gauge	`model`	Effective `max_inflight × healthy_replicas`; `0` means no count backstop

The global occupancy controller updates all four gauges when a request is admitted, released, adjusted to exact input usage, or grown by streamed output. The healthy-replica count is recomputed for each admission and floored at 1, so the budget and backstop gauges follow the live serving pool. Per-tenant RPM rejects also increment hivenet_router_tenant_rate_limited_total in addition to the unified b4_rpm reason below.

Rejection reasons

`reason`	Gate and response
`b1`	Per-request input-token or image cap; HTTP `400 input_too_long`
`b2`	Global occupancy budget or `max_inflight`; HTTP `429 concurrency_limit_exceeded`
`b3`	Front-door `shed_if` pressure gate; HTTP `429 concurrency_limit_exceeded`
`b4_occupancy`	Serverless per-key occupancy share; HTTP `429 rate_limit_exceeded`
`b4_itpm`	Serverless per-key input tokens per minute; HTTP `429 rate_limit_exceeded`
`b4_otpm`	Serverless per-key output tokens per minute; HTTP `429 rate_limit_exceeded`
`b4_rpm`	Requests per minute; HTTP `429 rate_limit_exceeded`

PromQL examples

Occupancy utilization

hivenet_router_admission_occupancy_tokens
  / hivenet_router_admission_budget_tokens

Exclude a zero or unset budget when using this expression in an alert:

(
  hivenet_router_admission_occupancy_tokens
    / hivenet_router_admission_budget_tokens
)
and on (model)
hivenet_router_admission_budget_tokens > 0

In-flight concurrency utilization

(
  hivenet_router_admission_inflight_requests
    / hivenet_router_admission_max_inflight
)
and on (model)
hivenet_router_admission_max_inflight > 0

Rejections by gate

sum by (reason, model) (
  rate(hivenet_router_admission_rejections_total[5m])
)

Front-door shed rate

sum by (model) (
  rate(hivenet_router_admission_rejections_total{reason="b3"}[5m])
)

Serverless per-key gate pressure

sum by (reason, model) (
  rate(hivenet_router_admission_rejections_total{reason=~"b4_.*"}[5m])
)

The rejection counter has no tenant_id or key_id label, so it can show that a per-key gate is firing but cannot identify the individual key. Correlate the time window with tenant failure and token-limit metrics or audit records.

Interpret the signals

Occupancy near budget with rising b2 means the model is at its global KV-occupancy limit.
b3 rising while occupancy remains below budget means live engine pressure is worse than the token estimate suggests, or the configured shed threshold is deliberately more conservative.
Sustained b4_itpm, b4_otpm, or b4_occupancy can indicate abusive traffic or undersized serverless key limits.
A nonzero max_inflight with high concurrency utilization can bind before the token budget does when requests are small.

See Admission control for gate order, reservation behavior, and client-facing errors.

Getting started

Deploy Hivenet Router

Use the API

Routing and policies

Security and auth

Observability

Integrations

Reference

Project

Metrics

Rejection reasons

PromQL examples

Occupancy utilization

In-flight concurrency utilization

Rejections by gate

Front-door shed rate

Serverless per-key gate pressure

Interpret the signals

​Metrics

​Rejection reasons

​PromQL examples

​Occupancy utilization

​In-flight concurrency utilization

​Rejections by gate

​Front-door shed rate

​Serverless per-key gate pressure

​Interpret the signals

Metrics

Rejection reasons

PromQL examples

Occupancy utilization

In-flight concurrency utilization

Rejections by gate

Front-door shed rate

Serverless per-key gate pressure

Interpret the signals