Implementation:Kserve Kserve TrainedModel Reconciler

Knowledge Sources	KServe KServe Multi-Model Serving
Domains	MLOps, Model_Management, Kubernetes
Last Updated	2026-02-13 00:00 GMT

Overview

Concrete Go reconciler that manages TrainedModel lifecycle including validation, ConfigMap synchronization, status tracking, and cascade deletion.

Description

The TrainedModelReconciler manages the full lifecycle of TrainedModel resources. It validates the parent ISVC, manages finalizers for cleanup, delegates to ModelConfigReconciler for ConfigMap writes, and updates status conditions via updateConditions() and updateStatus().

ConfigsDelta.Process() in configmap.go handles encoding/decoding of models.json in the ConfigMap data. The MemoryStrategy.GetOrAssignShard() always returns shard 0 (single shard mode).

Usage

This reconciler runs automatically in the KServe controller manager. It processes TrainedModel create/update/delete events.

Code Reference

Source Location

• Repository: kserve
• File: pkg/controller/v1alpha1/trainedmodel/controller.go, Lines 65-308
• File: pkg/controller/v1alpha1/trainedmodel/reconcilers/modelconfig/modelconfig_reconciler.go, Lines 52-94
• File: pkg/modelconfig/configmap.go, Lines 85-117

Signature

// TrainedModelReconciler reconciles TrainedModel resources
type TrainedModelReconciler struct {
    Client                client.Client
    Clientset             kubernetes.Interface
    Log                   logr.Logger
    Scheme                *runtime.Scheme
    Recorder              record.EventRecorder
    ModelConfigReconciler *modelconfig.ModelConfigReconciler
}

// Reconcile handles TrainedModel create/update/delete
func (r *TrainedModelReconciler) Reconcile(ctx context.Context,
    req ctrl.Request) (ctrl.Result, error)

// updateConditions checks ISVC ready, MMS enabled, memory available
func (r *TrainedModelReconciler) updateConditions(ctx context.Context,
    tm *v1alpha1.TrainedModel) error

// ConfigsDelta.Process encodes/decodes models.json in ConfigMap
func (cd *ConfigsDelta) Process(configMap *v1.ConfigMap) error

Import

import "github.com/kserve/kserve/pkg/controller/v1alpha1/trainedmodel"

I/O Contract

Inputs

Name	Type	Required	Description
req	ctrl.Request	Yes	Namespaced name of TrainedModel to reconcile

Outputs

Name	Type	Description
ConfigMap update	v1.ConfigMap	models.json updated with model spec
Status conditions	[]Condition	InferenceServiceReady, IsMMSPredictor, MemoryResourceAvailable
Status URL	URL	Per-model prediction URL derived from parent ISVC

Usage Examples

Monitor Reconciliation

# Check TrainedModel status conditions
kubectl get trainedmodel model1-sklearn -o yaml

# Expected conditions when healthy:
# conditions:
#   - type: InferenceServiceReady
#     status: "True"
#   - type: IsMMSPredictor
#     status: "True"
#   - type: MemoryResourceAvailable
#     status: "True"
#   - type: Ready
#     status: "True"

# Check ConfigMap content
kubectl get configmap modelconfig-sklearn-iris-example-0 -o yaml
# data:
#   models.json: '[{"modelName":"model1-sklearn",...}]'

Related Pages

Implements Principle

• Principle:Kserve_Kserve_Model_Lifecycle_Management

Requires Environment

• Environment:Kserve_Kserve_Kubernetes_Cluster