backendregistry

func NewKubernetesBackendRegistry ¶

func NewKubernetesBackendRegistry(
	ctx context.Context,
	namespace, group string,
	opts ...Option,
) (vmcp.BackendRegistry, server.Watcher, error)

NewKubernetesBackendRegistry builds a live, Kubernetes-populated backend registry for an embedder, hiding the pkg/vmcp/k8s watch substrate.

It bundles the registry+watcher wiring that pkg/vmcp/cli/serve.go performs by hand: it creates an empty DynamicRegistry, builds a k8s.BackendWatcher against it (in-cluster rest.Config by default; override with WithRESTConfig), and starts the watcher in a background goroutine bound to ctx. The watcher's initial informer sync populates the registry — so the registry starts empty, and the constructor needs only pkg/vmcp/k8s, not the static-discovery path's pkg/groups / pkg/container/runtime dependencies.

Parameters:

• ctx: bounds the watcher goroutine's lifetime; cancel it to stop the watcher.
• namespace: the Kubernetes namespace to watch for backend resources.
• group: the MCPGroup reference whose backends are tracked (parity with cli.Serve's vmcpCfg.Group).

Returns:

• vmcp.BackendRegistry: the live registry the watcher mutates, kept current as backends come and go. The narrow return type hides the mutators at compile time but the value is NOT immutable (a type assertion can reach them); embedders should treat it as read-only and let the watcher own all writes. Pass it to core.New (core.Config.BackendRegistry) and server.Serve (server.ServerConfig.BackendRegistry).
• server.Watcher: a readiness handle for the /readyz endpoint; pass it to server.ServerConfig.Watcher to gate readiness on informer cache sync.

Lifecycle and failure semantics (the watcher runs in a background goroutine, so the contract is not visible in the signature):

• The goroutine runs until ctx is cancelled. The caller MUST cancel ctx to release the goroutine and the watcher's informer caches/connections; there is no separate Stop/Close handle, so a caller that never cancels ctx leaks them.
• A successful return means the watcher was constructed and started, NOT that it has connected or synced. If watcher.Start later fails (e.g. the REST config points nowhere), the error is logged, not returned — the caller holds a registry that never populates. The returned server.Watcher is the only signal that distinguishes a watcher that started cleanly from one that failed, so the caller MUST wire it into server.ServerConfig.Watcher to gate /readyz on cache sync; otherwise a never-starting watcher goes undetected.