Observability

Observability

Everything is exposed through Spring Boot Actuator on the HTTP port (3000):

management.endpoints.web.exposure.include=health,info,prometheus,metrics

Micrometer with micrometer-registry-prometheus, so /actuator/prometheus and /actuator/metrics both work.

---

1. Bus meters

Wired by BusMetricsBinder (com.evento.server.bus.spring):

Meter	Meaning
evento.server.connections	Connected bundle nodes
evento.server.connections.available	Nodes able to receive
evento.server.correlations.pending	Outstanding server-initiated correlations
evento.server.forwarding.table.size	In-flight relayed requests
evento.server.forwarded{path=raw|reencoded}	Zero-copy versus re-encoded forward counts

forwarded{path=reencoded} rising for Netty-to-Netty traffic means the zero-copy path has been lost — that is a regression, and tests pin the contract.

Not yet covered: heartbeat lag, per-type processing-duration timers, consumer lag.

---

2. Capacity meters

Meter	Meaning
evento.server.bus.executor.pool.size	Current threads
evento.server.bus.executor.max	Configured maximum
evento.server.bus.executor.active	Threads currently executing
evento.server.bus.executor.queue.depth	The early warning
evento.server.bus.executor.saturated	The alerting signal — only moves when the pool is at max and the queue is full

Alert on any sustained increase in saturated. Full playbook: Throughput and Capacity.

---

3. Parallel-consumption meters

Wired by ConsumerMetricsRegistry (com.evento.server.bus.spring) and fed by a periodic ConsumerStatsMessage that bundles push on the admin notification channel (setConsumerStatsInterval, default 30 s).

Why push, not poll? The counters live in bundle-side objects. Polling per scrape would make one scrape a round trip per consumer.

Meter	Tags
evento.consumer.executor.{capacity,in.flight,admitted,rejected,completed,failed}	bundle,instance,executor
evento.consumer.async.{in.flight,submit.timeouts,transient.failures}	bundle,instance,consumer,component

evento.consumer.executor.rejected is the alerting signal — the executor reporting it is the bottleneck.

Meters are removed on BusEvent.NodeLeft, so rolling restarts do not leak a time series per dead instance. Bundles with no consumer executor push nothing at all.

---

4. Health

Endpoint	Contents
/actuator/health	BusHealthIndicator — bus bound port and node counts; DataSource health via Actuator's built-in indicator
/actuator/health/liveness · /readiness	Kubernetes probes (management.endpoint.health.probes.enabled=true)

management.endpoint.health.show-details=when_authorized, so details need authentication.

---

5. Performance store

PerformanceStoreService persists HandlerInvocationCountPerformance and HandlerServiceTimePerformance; PerformanceController serves them to the GUI.

Sampling is evento.performance.capture.rate (1 in the shipped properties, 0.1 in the compose file); retention is evento.telemetry.ttl days.

This is what powers the GUI's performance model and time series — and, per design decision 9, it is all the framework does about load. Scaling belongs to your orchestrator.

---

6. Structured logging

SLF4J, key=value on every bus lifecycle transition. Events worth knowing:

Log event	Level	Meaning
event=bus_business_executor_saturated	WARN	Pool at max and queue full. Rate-limited by business-executor-saturation-warn-interval
event=bundle_correlation_expired … byType={…}	WARN	Requests expired; byType names the drowning payload type
event=disconnect_superseded_skip	—	A superseded session's disconnect was correctly ignored (see Server Bus § 4)
event=listener_error … decode failed for BundleDiscoveryInfo	—	Discovery metadata was rejected — the bundle works but the dashboard shows nothing for it (see Wire Protocol § 8)

---

7. Tracing

TracingAgent's default is an honest no-op — telemetry and metrics only. For real distributed tracing, wire a custom agent or SentryTracingAgent via EventoBundle.Builder.setTracingAgent.

---

8. A minimal alerting set

Alert on	Why
evento.server.bus.executor.saturated increasing	Congestion collapse is starting
evento.consumer.executor.rejected increasing	A consumer executor is the bottleneck
evento.server.forwarded{path=reencoded} rising for bundle traffic	Zero-copy regression
WARN event=bundle_correlation_expired	Capacity exhaustion; byType tells you where
/actuator/health not UP	Bus or datasource down

---

See also

• Throughput and Capacity · Parallel Consumers · Troubleshooting