reconciler

pkg/controller/reconciler

Entry point to the reconciliation pipeline. Two components:

• Reconciler — fires immediately on every watched-resource / HTTP-resource change and on whole-index events. No reconciler-level debounce. Publishes ReconciliationTriggeredEvent.
• Coordinator — leader-only adapter that consumes ReconciliationTriggeredEvent and drives the synchronous render/validate pipeline via a PipelineExecutor.

Reconciler

import (
    "context"

    "gitlab.com/haproxy-haptic/haptic/pkg/controller/reconciler"
)

// No configuration — the reconciler fires immediately on every event.
r := reconciler.New(bus, logger)

go r.Start(ctx)

Triggering Rules

The Reconciler adds zero latency: every event it handles fires a reconciliation immediately. Coalescing of rapid changes is the per-watcher debounce window's job (default 2s, pkg/k8s/types.DefaultDebounceInterval; EndpointSlice watchers use debounceInterval: "0" so pod-IP rotations react instantly during rolling restarts). Reload throttling is the deployer's minDeploymentInterval (bypassed by the runtime-eligible fast path). This split keeps single ingress flips and rolling-restart endpoint rotations both fast without a reconciler-level refractory.

The initial-sync filter exists because ResourceIndexUpdatedEvent fires for every object as stores hydrate. Early reconciliations there would run against an incomplete store, so IndexSynchronizedEvent (which fires once every watcher finishes its initial list) is the correct first-reconciliation trigger.

Coordinator

coord := reconciler.NewCoordinator(&reconciler.CoordinatorConfig{
    EventBus:      bus,
    Pipeline:      pipeline,       // implements PipelineExecutor
    StoreProvider: storeProvider,  // pkg/stores.StoreProvider
    Logger:        logger,
})
go coord.Start(ctx)

Leader-only adapter around pkg/controller/pipeline:

1. Subscribe in Start (leader-only subscription pattern — not in the constructor, because followers must not subscribe).
2. On ReconciliationTriggeredEvent, publish ReconciliationStartedEvent.
3. Call Pipeline.Execute(ctx, storeProvider) synchronously — render + validate + build render context in one atomic step.
4. Publish the results: TemplateRenderedEvent + ValidationCompletedEvent on success, ReconciliationFailedEvent (with a PipelineError carrying the failing phase via errors.As) on error. Either path ends with ReconciliationCompletedEvent so the metrics adapter can close its histogram observation.

The pipeline is called directly, not through another event hop. That's deliberate: from the controller's perspective a reconciliation is one atomic stage, so making it a function call keeps error propagation straightforward and avoids inter-stage synchronization events.

See Also

• pkg/controller/pipeline — Pipeline.Execute implementation driven by this coordinator
• pkg/controller/renderer / validator — pure stages composed into the pipeline
• pkg/controller/deployer — downstream consumer of TemplateRenderedEvent + ValidationCompletedEvent
• pkg/controller/deployer — the DriftPreventionMonitor here is the actual DriftPreventionTriggeredEvent source (see drift_monitor.go); pkg/controller/timers is just a SafeTimer wrapper, not an event publisher
• pkg/controller/reconciler/CLAUDE.md — developer context (immediate-triggering design, leadership-transition patterns)

License

Apache-2.0 — see root LICENSE.