Principle:Kserve Kserve Model Lifecycle Management

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

Overview

A controller-based lifecycle management pattern for dynamically adding, updating, and removing models from a shared InferenceService.

Description

Model Lifecycle Management handles the full lifecycle of models in multi-model serving:

• Create: TrainedModel created → controller validates → writes to ConfigMap → agent loads model
• Update: TrainedModel storageUri changed → agent detects diff → unload old → load new
• Delete: TrainedModel deleted → finalizer cleans up ConfigMap → agent unloads model
• Cascade: Deleting the parent InferenceService triggers deletion of all associated TrainedModels

The controller uses Kubernetes finalizers to ensure cleanup is completed before resource removal.

Usage

Manage the model inventory by creating and deleting TrainedModel resources. The controller and agent handle all the underlying operations automatically.

Theoretical Basis

# Lifecycle management flow (NOT implementation code)
Create:
  1. Fetch TrainedModel spec
  2. Validate parent ISVC (exists, MMS, memory)
  3. Add "trainedmodel.finalizer"
  4. Write model spec to ConfigMap via ModelConfigReconciler
  5. Update TrainedModel status conditions

Delete:
  1. Finalizer intercepts deletion
  2. Remove model from ConfigMap
  3. Agent detects removal → unload model → delete files
  4. Remove finalizer → allow Kubernetes to delete resource

Cascade:
  - ISVC deletion triggers TrainedModel deletion
  - Controller detects ISVC not found → allows cleanup

Related Pages

Implemented By

• Implementation:Kserve_Kserve_TrainedModel_Reconciler