metrics.hpp File Reference

Always-on, backend-agnostic prom metrics for the conduit event system. More...

#include <prom/prom.hpp>
#include <memory>

Include dependency graph for metrics.hpp:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Namespaces
namespace	conduit
namespace	conduit::metrics

Functions
std::shared_ptr< prom::Scope > &	conduit::metrics::scope ()
	The shared conduit scope; applies the conduit_ prefix to every metric.
prom::Counter &	conduit::metrics::events_published ()
	Envelopes that entered the dispatch pipeline (one per Bus::publish).
prom::Counter &	conduit::metrics::events_dropped ()
	Envelopes dropped before listener fan-out (a middleware before_dispatch returned false, or a routing conflict).
prom::Counter &	conduit::metrics::listener_invocations ()
	Listener handler calls during fan-out (counts every matched listener, summed over all delivered envelopes).
prom::Counter &	conduit::metrics::listener_errors ()
	Listener handlers that threw during fan-out.
prom::Counter &	conduit::metrics::transport_errors ()
	Transport-level inbound failures, selected by the transport label ("redis", "amqp", "zmq", "mqtt", "nats", …).
prom::Gauge &	conduit::metrics::listeners ()
	Live listener subscriptions across all buses in the process.
prom::Histogram &	conduit::metrics::dispatch_seconds ()
	Wall-clock seconds from publish to the end of the synchronous dispatch pipeline.

Detailed Description

Always-on, backend-agnostic prom metrics for the conduit event system.

Every accessor returns a long-lived metric handle created on first call via a shared prom::Scope named "conduit" (metric prefix "conduit_"), cached in a static-local. Metrics are always-on and safe by default: until a host application installs a backend via prom::Registry::global()->set_adapter(...), every metric op is a noexcept no-op through prom's NullAdapter. This header never installs an adapter.

The recording sites live in bus.hpp: the Bus instruments its own publish / fan-out / error / subscription paths directly rather than through a metrics middleware. That keeps event-level telemetry on by default with no consumer API call, and — unlike a middleware — still counts envelopes published with flags::NoMiddleware, which bypass the middleware pipeline entirely. Pool / worker / thread telemetry comes for free from threadman's own prom instrumentation. Histograms observe seconds.

Labels. The bus attaches the most specific label it can derive at each site: event-level series carry {event=<name>} (the concrete event name, or a listener's name/pattern for conduit_listeners), and conduit_transport_errors_total carries {transport=<adapter>}. conduit emits whatever is derivable and leaves cardinality control to the consumer — a deployment that doesn't want per-event series drops the label at scrape/relabel time.