Skip to main content

Storage Options for Model Artifacts in KServe

KServe supports various storage options for serving machine learning models, allowing you to integrate with different storage systems based on your organizational needs, security requirements, and infrastructure setup. This document provides an overview of the supported storage providers and configurations in KServe.

Storage Providers​

KServe can serve models from various storage locations, including:

Cloud Storage​

Other Storage Options​

Storage Initializer​

KServe's storage initializer is responsible for downloading model artifacts from their source location to a local directory where the model server can access them. The storage initializer works as an init container that downloads the model and makes it available to the model server container.

Key capabilities of the storage initializer:

  • Downloads models from different storage providers
  • Handles authentication to secure storage systems
  • Supports custom certificates for secure connections
  • Provides options for direct volume mounting with PVCs
  • Enables modelcars for efficient model serving from OCI images

Security and Authentication​

When accessing storage providers that require authentication, KServe supports various authentication methods:

  • Self-Signed Certificates - For storage systems with self-signed SSL certificates
  • Service Account Credentials - For cloud storage providers
  • Access Keys and Secrets - For S3-compatible storage systems

Custom Storage Configuration​

For advanced use cases, KServe offers two main customization paths:

  1. Storage Containers - Define custom initialization logic using the ClusterStorageContainer CRD
  2. Storage Credentials - Configure storage credentials using Kubernetes secrets