Internals

Read from the source at commit 58d137d. Every claim here points at a file and line.

Code map

Path	Responsibility
cmd/manager/main.go	Controller-manager entry point; builds the Manager, registers reconcilers and webhooks (:99)
pkg/apis/serving/v1beta1	InferenceService, PredictorSpec, ComponentExtensionSpec, status types
pkg/apis/serving/v1alpha1	ServingRuntime, InferenceGraph, TrainedModel, LLMInferenceService
pkg/controller/v1beta1/inferenceservice	The isvc reconcile loop and component reconcilers
pkg/webhook/admission/pod	Pod mutating webhook and storage-initializer injection
pkg/constants/constants.go	Deployment modes and defaults (:547)
python/kserve/kserve	Python data-plane model servers (V1/V2 protocol)

Core data structures

• InferenceService / InferenceServiceSpec (pkg/apis/serving/v1beta1/inference_service.go:147, :24). Predictor PredictorSpec is required; Transformer and Explainer are optional pointers (:27, :31, :35). Status carries +kubebuilder:pruning:PreserveUnknownFields.
• ComponentExtensionSpec (pkg/apis/serving/v1beta1/component.go:82). The shared scaling and ops knobs for every component: MinReplicas *int32 where 0 means scale-to-zero (:85), MaxReplicas (:88), ScaleMetric (:98), and CanaryTrafficPercent (:114).
• ServingRuntimeSpec and SupportedModelFormat (pkg/apis/serving/v1alpha1/servingruntime_types.go:128, :31). SupportedModelFormats[].Name/Version/AutoSelect/Priority drive runtime auto-selection.
• InferenceServiceStatus (pkg/apis/serving/v1beta1/inference_service_status.go:34). Inlines Knative's duckv1.Status (:42) for conditions.
• DeploymentModeType (pkg/constants/constants.go:547). Standard / Knative / ModelMesh plus legacy aliases; the axis the whole reconcile branches on.

A path worth tracing

Deployment-mode resolution is the branch everything downstream depends on. ParseDeploymentMode normalizes legacy names before any reconcile decision:

// pkg/constants/constants.go:559
func ParseDeploymentMode(mode string) DeploymentModeType {
    if mode == "" {
        return DefaultDeployment
    }
    deploymentMode := DeploymentModeType(mode)
    switch deploymentMode {
    case LegacyRawDeployment:
        return Standard
    case LegacyServerless:
        return Knative
    default:
        return deploymentMode
    }
}

DefaultDeployment = Standard (constants.go:554), so an isvc with no mode annotation lands on plain Kubernetes. In the reconcile loop, the resolved mode reaches Predictor.Reconcile, which branches at predictor.go:204:

Predictor.Reconcile (predictor.go:85)
  -> if p.deploymentMode == constants.Standard (predictor.go:204)
       -> reconcileRawDeployment (predictor.go:744)
            -> raw.NewRawKubeReconciler  => Deployment + Service + HPA (predictor.go:771)
  -> else (Knative)
       -> knative.NewKsvcReconciler      => Knative Service (predictor.go:836)

Things that surprised me

The model never ships inside the server image. The pod mutating webhook does the heavy lifting: Mutator.Handle (pkg/webhook/admission/pod/mutator.go:46) calls InjectStorageInitializer (mutator.go:133), and StorageInitializerInjector.InjectStorageInitializer (pkg/webhook/admission/pod/storage_initializer_injector.go:668) appends an init container with params.PodSpec.InitContainers = append(...) (:441, :499). The init container downloads from storageUri into an emptyDir mounted at /mnt/models (:483, comment: Pvc will be mount to /mnt/models rather than /mnt/pvc.), which the model container then reads. The serving runtime stays model-agnostic.

The other surprise is the Knative legacy in the status type. Even in Standard mode the status inlines duckv1.Status (inference_service_status.go:42), a Knative condition duck type carried over from KFServing. The cross-component aggregation via PropagateCrossComponentStatus (controller.go:339-340) only fires in Knative mode.