Metrics

Prometheus metrics namespaced under runner_*. They exist only while a d3x-run process is running with -metrics-port set — a CLI one-shot (run-flow / run-suite) or another long-lived entrypoint that serves /metrics. There is no always-on daemon metrics port separate from the process you start.

Process-scoped

Metrics are not persisted. When the process exits, the endpoint and counters go with it. Scrape during the run (or keep a long-lived process) if you need samples.

Synopsis

Enable on a CLI run:

Bash

The flag is -metrics-port (also accepted as -metrics_port). Default is 0 (off). Serving metrics requires the obs.prometheus license capability.

Counters

The message and NGAP error counters instrument the NGAP pipeline only. SBI, REST, Diameter, and PFCP traffic is not counted here — track those flows through runner_flows_total and the latency histograms instead.

Name	Labels	Description
`runner_messages_total`	direction, msg_type	Total NGAP messages pipelined
`runner_message_bytes_total`	direction, msg_type	Total bytes of NGAP messages pipelined
`runner_ngap_encode_errors_total`	msg_type	NGAP encode failures inside EncoderStage
`runner_ngap_decode_errors_total`	msg_type	NGAP decode failures inside DecoderStage
`runner_flows_total`	flow, result	Completed flows (`result`: `success`, `failed`, `timeout`)
`runner_ue_deregistrations_total`	—	UE deregistrations observed (inner-NAS DeregistrationAccept)
`runner_procedure_failures_total`	procedure	Procedure failures observed on the wire
`runner_fsm_state_timeouts_total`	flow, state	`on_timeout` firings
`runner_bus_handler_total`	address	Total events dispatched per bus address
`runner_bus_handler_errors_total`	address	Handler-level errors (not business errors routed to internal.error)
`runner_bus_events_dropped_total`	reason	Events dropped (`reason`: `no_handler` for an unregistered address, `stopped` after the bus shuts down)

Gauges

Name	Labels	Description
`runner_flows_active`	flow	UEs whose flow is in-flight (between FSM start and terminal state)
`runner_gnb_registered`	—	gNBs that have completed NG Setup (incremented on NGSetupResponse)
`runner_ue_registered`	—	UEs currently registered with the 5GC
`runner_pdu_sessions_active`	—	PDU sessions currently active
`runner_bus_queue_depth`	—	Live event-bus queue depth

Histograms

Name	Labels	Description
`runner_pipeline_stage_seconds`	direction, stage, msg_type	Wall-clock time per pipeline stage per message
`runner_flow_latency_seconds`	flow, result	End-to-end flow latency, FSM start → terminal state
`runner_step_latency_seconds`	flow, from_state, to_state	FSM state dwell time per transition
`runner_bus_handler_duration_seconds`	address	Wall-clock time per event-bus handler invocation

Bucket layouts

runner_pipeline_stage_seconds and runner_step_latency_seconds use a stage-tuned bucket set covering 100 µs to 10 s in roughly half-decade steps:

Code

runner_flow_latency_seconds uses a flow-tuned set covering 5 ms to 60 s:

Code

runner_bus_handler_duration_seconds uses a finer low end (50 µs) for handler work:

Code

Examples

Live success rate per flow

promql

P95 flow latency

promql

Bus backpressure

promql

Notes

All custom series are namespaced runner_*.
result labels on counters and histograms are kept low-cardinality (small enum) so they're safe to slice and dice.
CLI one-shots register metrics on an isolated registry for the lifetime of that process; scrape while the run is in flight.
Agent mode does not expose a separate admin metrics port by itself — attach scrape configuration to whatever long-lived process you run with -metrics-port enabled, or use the control plane's own observability surface.

Where to go next

Running flows and suites — enable -metrics-port on a one-shot.
Agent mode — continuous workers attached to the control plane.
CI integration — gate on exit codes; scrape only when the job process is still alive.