Observability (OTEL)

pkg/resilience/otel provides an OpenTelemetry Plugin for the resilience package. Opt-in — the core package has zero external dependencies.

Usage

import rsotel "github.com/thumbrise/autosolve/pkg/resilience/otel"

client := resilience.NewClient(rsotel.Plugin())

That's it. All calls through this Client emit OTEL metrics automatically.

Metrics

Metric	Type	Description
resilience.call.total	Counter	Every fn call (including retries)
resilience.call.duration_seconds	Histogram	Duration of each fn call
resilience.call.errors	Counter	fn calls that returned an error
resilience.retry.total	Counter	Retry decisions (label: option)
resilience.retry.wait_seconds	Histogram	Backoff wait duration (label: option)

The option label carries the retry option name (e.g. "node", "service", "unregistered"). This enables per-category dashboards in Grafana.

How it works

The Plugin implements resilience.Plugin — it returns Events hooks:

• OnAfterCall — increments call.total, records call.duration_seconds, increments call.errors on failure
• OnBeforeWait — increments retry.total, records retry.wait_seconds with the option name label

No OnBeforeCall — the Plugin doesn't need it. Events are additive — multiple Plugins can coexist.

No OTEL SDK? No overhead.

If no OTEL SDK is configured, the default no-op meter produces no-op instruments. Zero allocations, zero overhead. The Plugin is always safe to register.

Design

OTEL is a Plugin, not an Option. It observes — doesn't control execution. Shared across all calls via Client. This is the canonical example of when to use Plugin vs Option.