e2e/

README

E2E Tests for Kthena

This directory contains end-to-end (E2E) tests for the Kthena project using Kind (Kubernetes in Docker).

Overview

The E2E tests use helm to install kthena into a Kind cluster and verify core functionality. Tests are organized into categories that can run in parallel on CI.

Test Categories

Category	Description	Make Target
controller-manager	ModelServing, ModelBooster, Autoscaling	make test-e2e-controller-manager
router	ModelRoute without Gateway API	make test-e2e-router
gateway-api	ModelRoute with Gateway API	make test-e2e-gateway-api
gateway-inference-extension	HTTPRoute, InferencePool	make test-e2e-gateway-inference-extension

Prerequisites

• Kind must be installed
• Go 1.24
• Docker (required by Kind)
• Helm (required to install helm charts)

Test Environment

The tests create a Kind cluster with the following characteristics:

• Cluster Name: kthena-e2e (can be overridden with CLUSTER_NAME env var)
• Kubernetes Version: v1.31.0
• Test Namespace: dev

Running the Tests

Run All Tests (Legacy)

make test-e2e
make test-e2e-cleanup

Run Specific Category

# Controller manager tests
make test-e2e-controller-manager

# Router tests (no Gateway API)
make test-e2e-router

# Gateway API tests
make test-e2e-gateway-api

# Gateway Inference Extension tests
make test-e2e-gateway-inference-extension

# Cleanup after any test
make test-e2e-cleanup

Environment Variables

• CLUSTER_NAME: Override Kind cluster name (default: kthena-e2e)
• TEST_CATEGORY: Specify which CRDs to install (used by setup.sh)
• HUB: Docker image registry (default: ghcr.io/volcano-sh)
• TAG: Docker image tag (default: latest)

Local Testing Considerations

CPU Limitations (AVX-512)

Some E2E tests (e.g., TestModelCR in model_booster_test.go) use vLLM-based images (ghcr.io/huntersman/vllm-cpu-env:latest).

Important: The official vLLM CPU backend relies heavily on the AVX-512 instruction set.

• Linux Servers: Most modern server-grade CPUs (Intel Xeon Scalable, AMD EPYC "Genoa") support AVX-512.
• Laptops/Intel Macs: Most consumer-grade Intel CPUs (especially older ones or 12th+ Gen Core where Intel disabled AVX-512) do not support it.
• Apple Silicon: Architecture mismatch (ARM64 vs x86_64).

If you run these tests on an incompatible machine, the vLLM containers may crash with SIGILL (Illegal Instruction). For more details, see the vLLM CPU installation guide.

Running Non-vLLM Tests Locally

Tests that do not require the vLLM image (like webhook validation/mutation tests) can be executed locally even without AVX-512 support.

For example, to run the controller manager webhook tests:

go test -v ./test/e2e/controller-manager/ -run TestWebhook