feat: CloudEvents ingest compliance and add OpenTelemetry distributed tracing

[kaio6fellipe]

Summary

Add CloudEvents compliance and OpenTelemetry distributed tracing to Argo Events.

CloudEvents ingest compliance (Closes #2011, Related: #1015, #983)

Argo Events currently discards all incoming CloudEvent metadata and regenerates it with internal values. This PR preserves the original CloudEvent attributes (id, source, type, subject, time) and extension attributes (e.g., traceparent, tracestate) when an incoming HTTP request is a valid CloudEvent.

• Add WithCloudEvent and individual attribute Option constructors
• Detect incoming CloudEvents via CE SDK HTTP binding in webhook handler
• Add Extensions field to EventContext for pipeline-wide extension propagation
• Preserve extensions in sensor convertEvent()

OpenTelemetry Distributed Tracing (Closes #1111)

Add opt-in distributed tracing across the event pipeline (EventSource → EventBus → Sensor → Trigger) using OpenTelemetry with W3C Trace Context propagation.

• New pkg/shared/tracing package with InitTracer, SpanFromCloudEvent, InjectTraceIntoCloudEvent
• eventsource.publish span wrapping event bus publish
• sensor.trigger span wrapping trigger execution with error recording
• Configuration via standard OTel env vars (OTEL_EXPORTER_OTLP_ENDPOINT) — fully opt-in, no-op when unset

Backward Compatibility

• Zero breaking changes — existing YAML resources work without modification
• All 30+ event sources that don't receive CloudEvents behave identically
• Tracing has zero performance impact when disabled
• EventContext.Extensions is optional (omitempty) with next available protobuf field number

Tests and behavior

The implementation was tested on a real GKE environment with Jaeger All-in-One as the trace backend. The test setup consists of three interconnected EventSource/Sensor pairs spanning two namespaces (sandbox and platform), exercising the full event pipeline including retries and dead-letter queue (DLQ) routing.

Test environment

• Argo Events deployed via Helm chart with custom image containing all changes
• Jaeger All-in-One deployed in observability namespace as OTLP collector + trace UI
• EventBus: Kafka (managed with Strimzi)
• OTel tracing enabled via OTEL_EXPORTER_OTLP_ENDPOINT and OTEL_SERVICE_NAME env vars on EventSource/Sensor pod templates

Test scenario

A single HTTP POST triggers a chain across 3 EventSources and 3 Sensors:

1. another-webhook (EventSource) → another-sensor (Sensor) — forwards via HTTP trigger
2. helloworld-webhook (EventSource) → helloworld-http (Sensor) — processor endpoint is down, retries 3x, then fires DLQ trigger
3. dlqueue-webhook (EventSource) → dlqueue-processor (Sensor) — processes the dead letter

graph TD
    A["<b>eventsource.publish</b><br/>another-webhook<br/>event.id: ba49ae95...<br/>47.0ms ✅"]
    B["<b>sensor.trigger</b><br/>another-sensor → another-sensor<br/>6.3ms ✅"]
    C["<b>eventsource.publish</b><br/>helloworld-webhook<br/>event.id: d14e2273...<br/>3.9ms ✅"]
    D["<b>sensor.trigger</b><br/>helloworld-http → helloworld-http<br/>31.9ms ❌ retry 1"]
    E["<b>sensor.trigger</b><br/>helloworld-http → helloworld-http<br/>0.8ms ❌ retry 2"]
    F["<b>sensor.trigger</b><br/>helloworld-http → helloworld-http<br/>0.7ms ❌ retry 3"]
    G["<b>sensor.trigger</b><br/>helloworld-http → dlq-http-trigger<br/>7.2ms ✅"]
    H["<b>eventsource.publish</b><br/>dlqueue-webhook<br/>event.id: 7ed81bf0...<br/>4.6ms ✅"]
    I["<b>sensor.trigger</b><br/>dlqueue-processor → dlqueue-processor<br/>2.6ms ✅"]

    A --> B
    B -->|"HTTP POST"| C
    C --> D
    C --> E
    C --> F
    C -->|"retries exhausted"| G
    G -->|"HTTP POST to DLQ"| H
    H --> I

    style A fill:#2d6a4f,color:#fff
    style B fill:#2d6a4f,color:#fff
    style C fill:#2d6a4f,color:#fff
    style D fill:#ae2012,color:#fff
    style E fill:#ae2012,color:#fff
    style F fill:#ae2012,color:#fff
    style G fill:#2d6a4f,color:#fff
    style H fill:#2d6a4f,color:#fff
    style I fill:#2d6a4f,color:#fff

Loading

Jaeger trace view

The entire chain appears as a single connected trace (a613bf5b2cccbb628b67bacb729a0e46) with 9 spans across 5 services:

Jaeger service map

The service dependency graph shows the flow between all instrumented services:

Argo Events resources

EventSource and Sensors:

Key observations

1. End-to-end trace propagation works — W3C traceparent is injected by the HTTP trigger into outgoing requests and extracted by downstream webhook EventSources, linking all hops into a single trace.

2. OTEL_SERVICE_NAME is respected — Each EventSource/Sensor appears with its configured service name in Jaeger (e.g., helloworld-webhook, dlqueue-processor) rather than the generic defaults.

3. Retries are visible — Failed trigger attempts appear as individual sensor.trigger spans with ERROR status, making it easy to identify retry behavior and eventual DLQ routing.

4. event.id changes at each EventSource — This is correct per the CloudEvents spec. Each EventSource is an independent event producer and generates a new ID. The trace context (traceparent) is what links the hops, not the event ID.

5. Zero impact when disabled — When OTEL_EXPORTER_OTLP_ENDPOINT is not set, the OTel tracer is a no-op with no performance overhead.

Checklist:

• My organization is added to USERS.md.