使用 Operator 安装
Semantic Router Operator 提供了一种 Kubernetes 原生的方式,通过自定义资源定义(CRD)来部署与管理 vLLM Semantic Router 实例。它可以在 Kubernetes 与 OpenShift 平台上简化部署、配置与生命周期管理。
特性
- 声明式部署:使用 Kubernetes CRD 定义语义路由实例
- 自动配置:生成并管理用于语义路由配置的 ConfigMap
- 持久化存储:管理用于 ML 模型存储的 PVC,并自动处理生命周期
- 平台探测:自动识别 OpenShift 或标准 Kubernetes,并做相应配置
- 内置可观测性:默认支持指标、链路追踪与监控
- 生产能力:HPA、Ingress、Service Mesh 集成、Pod Disruption Budget
- 默认安全:移除全部 capability,禁止特权提升
前置条件
- Kubernetes 1.24+ 或 OpenShift 4.12+
- 已配置好的
kubectl或oc命令行 - 集群管理员权限(用于安装 CRD)
安装
选项 1:使用 Kustomize(标准 Kubernetes)
# Clone the repository
git clone https://github.com/vllm-project/semantic-router
cd semantic-router/deploy/operator
# Install CRDs
make install
# Deploy the operator
make deploy IMG=ghcr.io/vllm-project/semantic-router/operator:latest
验证 operator 正在运行:
kubectl get pods -n semantic-router-operator-system
选项 2:使用 OLM(OpenShift)
适用于通过 Operator Lifecycle Manager 部署到 OpenShift 的场景:
cd semantic-router/deploy/operator
# Build and push to your registry (Quay, internal registry, etc.)
podman login quay.io
make podman-build IMG=quay.io/<your-org>/semantic-router-operator:latest
make podman-push IMG=quay.io/<your-org>/semantic-router-operator:latest
# Deploy using OLM
make openshift-deploy
部署你的第一个 Router
使用示例配置快速开始
根据你的基础设施选择一个预配置示例:
# Simple standalone deployment with KServe backend
kubectl apply -f https://raw.githubusercontent.com/vllm-project/semantic-router/main/deploy/operator/config/samples/vllm.ai_v1alpha1_semanticrouter_simple.yaml
# Full-featured OpenShift deployment with Routes
kubectl apply -f https://raw.githubusercontent.com/vllm-project/semantic-router/main/deploy/operator/config/samples/vllm.ai_v1alpha1_semanticrouter_openshift.yaml
# Gateway integration mode (Istio/Envoy Gateway)
kubectl apply -f https://raw.githubusercontent.com/vllm-project/semantic-router/main/deploy/operator/config/samples/vllm.ai_v1alpha1_semanticrouter_gateway.yaml
# Llama Stack backend discovery
kubectl apply -f https://raw.githubusercontent.com/vllm-project/semantic-router/main/deploy/operator/config/samples/vllm.ai_v1alpha1_semanticrouter_llamastack.yaml
# Redis cache backend for production caching
kubectl apply -f https://raw.githubusercontent.com/vllm-project/semantic-router/main/deploy/operator/config/samples/vllm.ai_v1alpha1_semanticrouter_redis_cache.yaml
# Milvus cache backend for large-scale deployments
kubectl apply -f https://raw.githubusercontent.com/vllm-project/semantic-router/main/deploy/operator/config/samples/vllm.ai_v1alpha1_semanticrouter_milvus_cache.yaml
# Hybrid cache backend for optimal performance
kubectl apply -f https://raw.githubusercontent.com/vllm-project/semantic-router/main/deploy/operator/config/samples/vllm.ai_v1alpha1_semanticrouter_hybrid_cache.yaml
# mmBERT 2D Matryoshka embeddings with layer early exit
kubectl apply -f https://raw.githubusercontent.com/vllm-project/semantic-router/main/deploy/operator/config/samples/vllm.ai_v1alpha1_semanticrouter_mmbert.yaml
# Complexity-aware routing for intelligent model selection
kubectl apply -f https://raw.githubusercontent.com/vllm-project/semantic-router/main/deploy/operator/config/samples/vllm.ai_v1alpha1_semanticrouter_complexity.yaml
自定义配置
创建一个 my-router.yaml 文件:
apiVersion: vllm.ai/v1alpha1
kind: SemanticRouter
metadata:
name: my-router
namespace: default
spec:
replicas: 2
image:
repository: ghcr.io/vllm-project/semantic-router/extproc
tag: latest
# Configure vLLM backend endpoints
vllmEndpoints:
# KServe InferenceService (RHOAI 3.x)
- name: llama3-8b-endpoint
model: llama3-8b
reasoningFamily: qwen3
loras:
- name: computer-science-expert
description: Adapter for advanced computer science prompts
backend:
type: kserve
inferenceServiceName: llama-3-8b
weight: 1
resources:
limits:
memory: "7Gi"
cpu: "2"
requests:
memory: "3Gi"
cpu: "1"
persistence:
enabled: true
size: 10Gi
storageClassName: "standard"
config:
providers:
defaults:
default_model: llama3-8b
default_reasoning_effort: medium
reasoning_families:
qwen3:
type: chat_template_kwargs
parameter: enable_thinking
models:
- name: llama3-8b
provider_model_id: llama3-8b
backend_refs:
- name: llama3-8b-endpoint
endpoint: llama-3-8b-predictor.default.svc.cluster.local:80
protocol: http
routing:
modelCards:
- name: llama3-8b
modality: text
capabilities: ["chat", "reasoning"]
decisions:
- name: default-route
description: Catch-all route
priority: 100
rules:
operator: AND
conditions: []
modelRefs:
- model: llama3-8b
use_reasoning: false
global:
stores:
semantic_cache:
enabled: true
backend_type: memory
max_entries: 1000
ttl_seconds: 3600
integrations:
tools:
enabled: true
top_k: 3
similarity_threshold: 0.2
model_catalog:
system:
prompt_guard: models/mmbert32k-jailbreak-detector-merged
modules:
prompt_guard:
enabled: true
model_ref: prompt_guard
threshold: 0.7
toolsDb:
- tool:
type: "function"
function:
name: "get_weather"
description: "Get weather information for a location"
parameters:
type: "object"
properties:
location:
type: "string"
description: "City and state, e.g. San Francisco, CA"
required: ["location"]
description: "Weather information tool"
category: "weather"
tags: ["weather", "temperature"]
应用配置:
kubectl apply -f my-router.yaml
spec.config 应使用与本地 config.yaml 相同的规范化 providers/routing/global 布局。spec.vllmEndpoints 仍是 Kubernetes 适配层,用于发现后端与 served-model alias;operator 在渲染 runtime config 时,会将其转换为规范化的 providers.models[].backend_refs[] 与 routing.modelCards 条目(包含可选的 loras)。
高级特性
Embedding 模型配置
operator 支持三种高性能 embedding 模型,用于语义理解与缓存。你可以根据场景配置这些模型以优化效果。
可用的 embedding 模型
-
Qwen3-Embedding(1024 维,32K 上下文)
- 适合:高质量语义理解与长上下文
- 场景:复杂查询、研究文档、细致分析
-
EmbeddingGemma(768 维,8K 上下文)
- 适合:更快性能与较好精度
- 场景:实时应用、高吞吐
-
mmBERT 2D Matryoshka(64-768 维,多语言)
- 适合:可通过 layer early exit 自适应权衡速度与质量
- 场景:多语言部署、需要灵活的质量/速度权衡
示例:mmBERT + Layer Early Exit
spec:
config:
global:
model_catalog:
embeddings:
semantic:
mmbert_model_path: "models/mom-embedding-ultra"
use_cpu: true
embedding_config:
model_type: "mmbert"
# Layer early exit: balance speed vs accuracy
# Layer 3: ~7x speedup (fast, good for high-volume queries)
# Layer 6: ~3.6x speedup (balanced - recommended)
# Layer 11: ~2x speedup (higher accuracy)
# Layer 22: full model (maximum accuracy)
target_layer: 6
# Dimension reduction for faster similarity search
# Options: 64, 128, 256, 512, 768
target_dimension: 256
preload_embeddings: true
enable_soft_matching: true
top_k: 1
min_score_threshold: "0.5"
stores:
semantic_cache:
enabled: true
backend_type: "memory"
embedding_model: "mmbert"
similarity_threshold: "0.85"
max_entries: 5000
ttl_seconds: 7200
完整示例可参考 mmbert sample configuration。
示例:Qwen3 + Redis Cache
spec:
config:
global:
model_catalog:
embeddings:
semantic:
qwen3_model_path: "models/qwen3-embedding"
use_cpu: true
stores:
semantic_cache:
enabled: true
backend_type: "redis"
embedding_model: "qwen3"
redis:
connection:
host: redis.cache-backends.svc.cluster.local
port: 6379
index:
vector_field:
dimension: 1024 # Qwen3 dimension
完整示例可参考 redis cache sample configuration。
基于复杂度的路由(Complexity-Aware Routing)
根据复杂度分类将请求路由到不同模型:简单请求走更快的模型,复杂请求走更强的模型。
示例配置
spec:
# Configure multiple backends with different capabilities
vllmEndpoints:
- name: llama-8b-fast
model: llama3-8b
reasoningFamily: qwen3
backend:
type: kserve
inferenceServiceName: llama-3-8b
weight: 2 # Prefer for simple queries
- name: llama-70b-reasoning
model: llama3-70b
reasoningFamily: deepseek
backend:
type: kserve
inferenceServiceName: llama-3-70b
weight: 1 # Use for complex queries
config:
# Define complexity rules
complexity_rules:
# Rule 1: Code complexity
- name: "code-complexity"
description: "Classify coding tasks by complexity"
threshold: "0.3" # Lower threshold works better for embedding-based similarity
# Examples of complex coding tasks
hard:
candidates:
- "Implement a distributed lock manager with leader election"
- "Design a database migration system with rollback support"
- "Create a compiler optimization pass for loop unrolling"
# Examples of simple coding tasks
easy:
candidates:
- "Write a function to reverse a string"
- "Create a class to represent a rectangle"
- "Implement a simple counter with increment/decrement"
# Rule 2: Reasoning complexity
- name: "reasoning-complexity"
description: "Classify reasoning and problem-solving tasks"
threshold: "0.3" # Lower threshold works better for embedding-based similarity
hard:
candidates:
- "Analyze the geopolitical implications of renewable energy adoption"
- "Evaluate the ethical considerations of AI in healthcare"
- "Design a multi-stage marketing strategy for a new product launch"
easy:
candidates:
- "What is the capital of France?"
- "How many days are in a week?"
- "Name three common pets"
# Rule 3: Domain-specific complexity with conditional application
- name: "medical-complexity"
description: "Classify medical queries (only for medical domain)"
threshold: "0.3" # Lower threshold works better for embedding-based similarity
hard:
candidates:
- "Differential diagnosis for chest pain with dyspnea"
- "Treatment protocol for multi-drug resistant tuberculosis"
easy:
candidates:
- "What is the normal body temperature?"
- "What are common symptoms of a cold?"
# Only apply this rule if domain signal indicates medical domain
composer:
operator: "AND"
conditions:
- type: "domain"
name: "medical"
工作原理:
- 输入查询会与
hard与easy的候选示例做相似度比较 - 相似度分数用于判定复杂度
- 输出 signals:
{rule-name}:hard、{rule-name}:easy或{rule-name}:medium - Router 根据 signals 选择后端模型
composer支持根据其他 signals 做条件性规则应用
完整示例可参考 complexity routing sample configuration。
验证部署
# Check the SemanticRouter resource
kubectl get semanticrouter my-router
# Check created resources
kubectl get deployment,service,configmap -l app.kubernetes.io/instance=my-router
# View status
kubectl describe semanticrouter my-router
# View logs
kubectl logs -f deployment/my-router
期望输出示例:
NAME PHASE REPLICAS READY AGE
semanticrouter.vllm.ai/my-router Running 2 2 5m
后端发现类型(Backend Discovery Types)
operator 支持三种后端发现方式,用于连接 semantic router 与 vLLM 模型服务。选择与你的基础设施匹配的类型 即可。
KServe InferenceService 发现
适用于 RHOAI 3.x 或独立 KServe 部署。operator 会自动发现 KServe 创建的 predictor service。
spec:
vllmEndpoints:
- name: llama3-8b-endpoint
model: llama3-8b
reasoningFamily: qwen3
backend:
type: kserve
inferenceServiceName: llama-3-8b # InferenceService in same namespace
weight: 1
适用场景:
- 运行在 Red Hat OpenShift AI(RHOAI)3.x
- 使用 KServe 做模型服务
- 需要自动服务发现
工作原理:
- 发现 predictor service:
{inferenceServiceName}-predictor - 使用 8443 端口(KServe 默认 HTTPS 端口)
- 与
SemanticRouter在同一命名空间内工作
Llama Stack Service 发现
通过 Kubernetes label selector 发现 Llama Stack 部署。
spec:
vllmEndpoints:
- name: llama-405b-endpoint
model: llama-3.3-70b-instruct
reasoningFamily: gpt
backend:
type: llamastack
discoveryLabels:
app: llama-stack
model: llama-3.3-70b
weight: 1
适用场景:
- 使用 Meta 的 Llama Stack 做模型服务
- 同时有多个 Llama Stack 服务、不同模型
- 需要基于 label 做服务发现
工作原理:
- 列出匹配 label selector 的 services
- 若匹配多个,默认使用第一个
- 从 service 定义中提取端口
直连 Kubernetes Service
可直连任意 Kubernetes service(vLLM、TGI 等)。
spec:
vllmEndpoints:
- name: custom-vllm-endpoint
model: deepseek-r1-distill-qwen-7b
reasoningFamily: deepseek
backend:
type: service
service:
name: vllm-deepseek
namespace: vllm-serving # Can reference service in another namespace
port: 8000
weight: 1
适用场景:
- 直接部署的 vLLM 服务
- 自定义模型服务器(OpenAI 兼容 API)
- 跨命名空间引用 service
- 希望完全控制 service endpoint
工作原理:
- 直接连接指定 service
- 不做 discovery,完全按显式配置
- 支持跨命名空间引用
多后端(Multiple Backends)
你可以配置多个后端,并用权重做负载均衡:
spec:
vllmEndpoints:
# KServe backend
- name: llama3-8b
model: llama3-8b
reasoningFamily: qwen3
backend:
type: kserve
inferenceServiceName: llama-3-8b
weight: 2 # Higher weight = more traffic
# Direct service backend
- name: qwen-7b
model: qwen2.5-7b
reasoningFamily: qwen3
backend:
type: service
service:
name: vllm-qwen
port: 8000
weight: 1
部署模式(Deployment Modes)
operator 支持两种部署模式,对应不同架构。
Standalone 模式(默认)
部署 semantic router,并带一个 Envoy sidecar 容器作为入口网关。
架构:
Client → Service (8080) → Envoy Sidecar → ExtProc gRPC → Semantic Router → vLLM
适用场景:
- 没有现成 service mesh 的简单部署
- 测试与开发
- 自包含部署、依赖最少
配置:
spec:
# No gateway configuration - defaults to standalone mode
service:
type: ClusterIP
api:
port: 8080 # Client traffic enters here
targetPort: 8080 # Envoy ingress port
grpc:
port: 50051 # ExtProc communication
targetPort: 50051
operator 行为:
- 在 pod spec 中部署两个容器:semantic router + Envoy sidecar
- Envoy 负责 ingress,并通过 ExtProc gRPC 转发到 semantic router
- status 显示
gatewayMode: "standalone"
Gateway Integration 模式
复用 已有 Gateway(Istio、Envoy Gateway 等),并要求你自行管理匹配的 HTTPRoute。
当前状态:controller 能解析所引用的 Gateway 并切换到 gateway mode,但自动创建 HTTPRoute 仍是占位实现。
架构:
Client → Gateway (Istio/Envoy) → user-managed HTTPRoute → Service (8080) → Semantic Router API → vLLM
适用场景:
- 已有 Istio 或 Envoy Gateway 部署
- 统一的 ingress 管理
- 共享网关的多租户场景
- 高级流量治理(熔断、重试、限流等)
配置:
spec:
gateway:
existingRef:
name: istio-ingressgateway # Or your Envoy Gateway name
namespace: istio-system
# Service only needs API port in gateway mode
service:
type: ClusterIP
api:
port: 8080
targetPort: 8080
operator 行为:
- 解析 referenced Gateway 并进入 gateway integration 模式
- 目前不创建 HTTPRoute,你需要自行 apply 与管理该资源
- pod spec 中不再部署 Envoy sidecar
- 设置
status.gatewayMode: "gateway-integration" - semantic router 以纯 API 模式运行(不启用 ExtProc)
示例: 参考 vllm.ai_v1alpha1_semanticrouter_gateway.yaml。
该示例仅为 gateway mode 配置 SemanticRouter 资源,并不会替你安装 Gateway 或 HTTPRoute。
OpenShift Routes
在 OpenShift 上,operator 可创建 Route 用于外部访问,并支持 TLS 终止。
基础 OpenShift Route
spec:
openshift:
routes:
enabled: true
hostname: semantic-router.apps.openshift.example.com # Optional - auto-generated if omitted
tls:
termination: edge # TLS terminates at Route, plain HTTP to backend
insecureEdgeTerminationPolicy: Redirect # Redirect HTTP to HTTPS
TLS 终止方式
- edge(推荐):TLS 终止在 Route,后端走明文 HTTP
- passthrough:TLS 透传到后端(要求后端支持 TLS)
- reencrypt:TLS 终止在 Route,并对后端重新加密
何时使用 OpenShift Routes
- 运行在 OpenShift 4.x
- 希望无需配置 Ingress 就提供外部访问
- 希望自动生成 hostname
- 需要 OpenShift 原生的 TLS 管理能力
状态信息
创建 Route 后可查看状态:
kubectl get semanticrouter my-router -o jsonpath='{.status.openshiftFeatures}'
输出示例:
{
"routesEnabled": true,
"routeHostname": "semantic-router-default.apps.openshift.example.com"
}
示例: 参考 vllm.ai_v1alpha1_semanticrouter_route.yaml。
如何选择配置
使用下面的决策树选择合适的配置:
┌─ Need to run on OpenShift?
│ ├─ YES → Use openshift sample (Routes + KServe/service backends)
│ └─ NO ↓
│
├─ Have existing Gateway (Istio/Envoy)?
│ ├─ YES → Use gateway sample (Gateway integration mode)
│ └─ NO ↓
│
├─ Using Meta Llama Stack?
│ ├─ YES → Use llamastack sample
│ └─ NO ↓
│
└─ Simple deployment → Use simple sample (standalone mode)
后端选择:
┌─ Running RHOAI 3.x or KServe?
│ ├─ YES → Use KServe backend type
│ └─ NO ↓
│
├─ Using Meta Llama Stack?
│ ├─ YES → Use llamastack backend type
│ └─ NO ↓
│
└─ Have direct vLLM service? → Use service backend type
架构
operator 会为每个 SemanticRouter 管理一整套资源:
┌─────────────────────────────────────────────────────┐
│ SemanticRouter CR │
│ apiVersion: vllm.ai/v1alpha1 │
│ kind: SemanticRouter │
└──────────────────┬──────────────────────────────────┘
│
▼
┌─────────────────────┐
│ Operator Controller │
│ - Watches CR │
│ - Reconciles state │
│ - Platform detection│
└─────────┬────────────┘
│
┌────────────┼────────────┬──────────────┐
▼ ▼ ▼ ▼
┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐
│Deployment│ │ Service │ │ConfigMap│ │ PVC │
│ │ │ - gRPC │ │ - config│ │ - models│
│ │ │ - API │ │ - tools │ │ │
│ │ │ - metrics│ │ │ │ │
└─────────┘ └─────────┘ └─────────┘ └─────────┘
管理的资源:
- Deployment:运行 semantic router pods,可配置副本数
- Service:暴露 gRPC(50051)、HTTP API(8080)与 metrics(9190)
- ConfigMap:包含 semantic router 配置与 tools database
- ServiceAccount:RBAC(可选,仅当指定时创建)
- PersistentVolumeClaim:ML 模型存储(可选,仅当启用 persistence 时创建)
- HorizontalPodAutoscaler:自动伸缩(可选,仅当启用 autoscaling 时创建)
- Ingress:外部访问(可选,仅当启用 ingress 时创建)
平台探测与安全
operator 会自动识别平台,并设置适配的安全上下文。
OpenShift 平台
当运行在 OpenShift 时,operator 会:
- 探测:检查是否存在
route.openshift.ioAPI 资源 - 安全上下文:不会设置
runAsUser、runAsGroup、fsGroup - 原因:让 OpenShift SCC 从 namespace 允许的 UID/GID 范围里分配
- 兼容性:
restrictedSCC(默认)与自定义 SCC
标准 Kubernetes
当运行在标准 Kubernetes 时,operator 会:
- 安全上下文:设置
runAsUser: 1000、fsGroup: 1000、runAsNonRoot: true - 原因:提供安全的默认值,满足常见的 Pod 安全标准
两个平台通用
无论平台如何:
- 移除全部 capability(
drop: [ALL]) - 禁止特权提升(
allowPrivilegeEscalation: false) - 除默认权限外不需要额外权限或 SCC
覆盖安全上下文
你也可以在 CR 中覆盖自动安全上下文:
spec:
# Container security context
securityContext:
runAsNonRoot: true
runAsUser: 2000
allowPrivilegeEscalation: false
capabilities:
drop:
- ALL
# Pod security context
podSecurityContext:
runAsNonRoot: true
runAsUser: 2000
fsGroup: 2000
在 OpenShift 上,建议省略 runAsUser 与 fsGroup,让 SCC 自动分配 UID/GID。
配置参考
镜像配置
spec:
image:
repository: ghcr.io/vllm-project/semantic-router/extproc
tag: latest
pullPolicy: IfNotPresent
imageRegistry: "" # 可选:自定义镜像仓库前缀
# 可选:镜像拉取 Secret
imagePullSecrets:
- name: ghcr-secret
Service 配置
spec:
service:
type: ClusterIP # 或 NodePort、LoadBalancer
grpc:
port: 50051
targetPort: 50051
api:
port: 8080
targetPort: 8080
metrics:
enabled: true
port: 9190
targetPort: 9190
持久化配置
spec:
persistence:
enabled: true
storageClassName: "standard" # 根据你的集群调整
accessMode: ReadWriteOnce
size: 10Gi
# 可选:使用现有 PVC
existingClaim: "my-existing-pvc"
# 可选:PVC annotation
annotations:
backup.velero.io/backup-volumes: "models"
operator 会在创建 PVC 之前验证指定的 StorageClass 是否存在。如果省略 storageClassName,则使用集群的默认 StorageClass。
StorageClass 示例:
- AWS EKS:
gp3-csi、gp2 - GKE:
standard、premium-rwo - Azure AKS:
managed、managed-premium - OpenShift:
gp3-csi、thin、ocs-storagecluster-ceph-rbd
语义缓存后端
operator 支持多种语义缓存后端。通过缓存相似查询及其响应,可以显著降低延迟和 token 用量。
operator 不会部署 Redis 或 Milvus。在将其用作缓存后端之前,你必须在集群中单独部署这些服务。operator 只负责配置 SemanticRouter,使其连接到现有的 Redis/Milvus 部署。
替 代方案: 如果希望自动部署 Redis/Milvus,可以考虑使用 Helm chart,它可以将缓存后端作为 Helm chart 依赖项进行部署。
支持的后端
1. 内存缓存(默认)
简单的内存缓存,适用于开发环境和小规模部署。
特性:
- 无外部依赖
- 访问速度快
- 不持久化(重启时清空)
- 受 pod 内存限制
配置:
spec:
config:
global:
stores:
semantic_cache:
enabled: true
backend_type: memory
similarity_threshold: "0.8"
max_entries: 1000
ttl_seconds: 3600
eviction_policy: fifo # fifo、lru 或 lfu
适用场景:
- 开发与测试
- 小规模部署(缓存查询少于 1000 条)
- 没有持久化要求
2. Redis 缓存
使用具备向量搜索能力的 Redis 构建高性能分布式缓存。
特性:
- 分布式且可扩展
- 持久化存储(使用 AOF/RDB)
- HNSW 或 FLAT 索引
- 生态支持广泛
前置条件:
- Redis 7.0+,并安装 RediSearch 模块
- 创建用于保存密码的 Kubernetes Secret:
kubectl create secret generic redis-credentials \
--from-literal=password='your-redis-password'
配置:
spec:
config:
global:
stores:
semantic_cache:
enabled: true
backend_type: redis
similarity_threshold: "0.85"
ttl_seconds: 3600
embedding_model: mmbert
redis:
connection:
host: redis.default.svc.cluster.local
port: 6379
database: 0
password_secret_ref:
name: redis-credentials
key: password
timeout: 30
tls:
enabled: false
index:
name: semantic_cache_idx
prefix: "cache:"
vector_field:
name: embedding
dimension: 768 # 与默认的 mmBERT ultra embedding 模型一致
metric_type: COSINE
index_type: HNSW
params:
M: 16
efConstruction: 64
search:
topk: 1
development:
auto_create_index: true
verbose_errors: true
适用场景:
- 中等规模的生产部署
- 需要持久化和高可用性
- 已有 Redis 基础设施
- 需要高速的内存性能
示例: vllm.ai_v1alpha1_semanticrouter_redis_cache.yaml
3. Milvus 缓存
面向大容量缓存生产部署的企业级向量数据库。
特性:
- 高度可扩展且支持分布式部署
- 高级索引(HNSW、IVF 等)
- 内置数据生命周期管理
- 支持高可用性
前置条件:
- Milvus 2.3+(standalone 或集群模式)
- 创建用于保存凭据的 Kubernetes Secret:
kubectl create secret generic milvus-credentials \
--from-literal=password='your-milvus-password'
配置:
spec:
config:
global:
stores:
semantic_cache:
enabled: true
backend_type: milvus
similarity_threshold: "0.90"
ttl_seconds: 7200
embedding_model: mmbert
milvus:
connection:
host: milvus-standalone.default.svc.cluster.local
port: 19530
database: semantic_router_cache
timeout: 30
auth:
enabled: true
username: root
password_secret_ref:
name: milvus-credentials
key: password
collection:
name: semantic_cache
description: "Semantic cache for LLM responses"
vector_field:
name: embedding
dimension: 768 # 与默认的 mmBERT ultra embedding 模型一致
metric_type: IP
index:
type: HNSW
params:
M: 16
efConstruction: 64
search:
params:
ef: 64
topk: 10
consistency_level: Session
performance:
connection_pool:
max_connections: 10
max_idle_connections: 5
batch:
insert_batch_size: 100
data_management:
ttl:
enabled: true
timestamp_field: created_at
cleanup_interval: 3600
development:
auto_create_collection: true
适用场景:
- 大规模生产部署
- 需要高级向量搜索能力
- 需要数据生命周期管理(TTL、compaction)
- 对高可用性和可扩展性有要求
示例: vllm.ai_v1alpha1_semanticrouter_milvus_cache.yaml
4. 混合缓存
将内存 HNSW 索引与 Milvus 持久化存储结合,以兼顾最佳性能和持久性。
特性:
- 使用 HNSW 进行高速内存搜索
- 在 Milvus 中持久化存储
- 兼具两者优势
- 自动同步
配置:
spec:
config:
global:
stores:
semantic_cache:
enabled: true
backend_type: hybrid
similarity_threshold: "0.85"
ttl_seconds: 3600
max_entries: 5000
eviction_policy: lru
embedding_model: mmbert
# HNSW 内存配置
hnsw:
use_hnsw: true
hnsw_m: 32
hnsw_ef_construction: 128
max_memory_entries: 5000
# Milvus 持久化存储(配置与 milvus 后端相同)
milvus:
connection:
host: milvus-standalone.default.svc.cluster.local
port: 19530
# ... 其余 milvus 配置
适用场景:
- 需要尽可能快的缓存查询
- 需要持久化和数据耐久性
- 愿意用内存换取性能
- 高吞吐量生产部署
示例: vllm.ai_v1alpha1_semanticrouter_hybrid_cache.yaml
部署 Redis
在使用 Redis 缓存后端之前,请先在集群中部署带有 RediSearch 模块的 Redis:
---
apiVersion: v1
kind: Namespace
metadata:
name: cache-backends
---
apiVersion: apps/v1
kind: Deployment
metadata:
name: redis
namespace: cache-backends
spec:
replicas: 1
selector:
matchLabels:
app: redis
template:
metadata:
labels:
app: redis
spec:
containers:
- name: redis
image: redis/redis-stack-server:latest
ports:
- containerPort: 6379
resources:
requests:
cpu: 100m
memory: 256Mi
limits:
cpu: 500m
memory: 512Mi
---
apiVersion: v1
kind: Service
metadata:
name: redis
namespace: cache-backends
spec:
type: ClusterIP
ports:
- port: 6379
targetPort: 6379
selector:
app: redis
应用配置:
kubectl apply -f redis-deployment.yaml
创建凭据 Secret:
kubectl create secret generic redis-credentials \
--from-literal=password='' # 不使用密码时留空,也可以设置你的密码
对于生产部署,建议使用:
- Redis Operator
- Redis Enterprise Operator
- 托管 Redis 服务(AWS ElastiCache、Azure Cache for Redis、GCP Memorystore)
部署 Milvus
在使用 Milvus 缓存后端之前,请先在集群中部署 Milvus:
---
apiVersion: apps/v1
kind: Deployment
metadata:
name: milvus-standalone
namespace: cache-backends
spec:
replicas: 1
selector:
matchLabels:
app: milvus
template:
metadata:
labels:
app: milvus
spec:
containers:
- name: milvus
image: milvusdb/milvus:v2.4.0
command: ["milvus", "run", "standalone"]
ports:
- containerPort: 19530
name: grpc
- containerPort: 9091
name: metrics
env:
- name: ETCD_USE_EMBED
value: "true"
- name: COMMON_STORAGETYPE
value: "local"
resources:
requests:
cpu: 500m
memory: 1Gi
limits:
cpu: 2000m
memory: 4Gi
volumeMounts:
- name: milvus-data
mountPath: /var/lib/milvus
volumes:
- name: milvus-data
emptyDir: {} # 生产环境请使用 PVC
---
apiVersion: v1
kind: Service
metadata:
name: milvus-standalone
namespace: cache-backends
spec:
type: ClusterIP
ports:
- port: 19530
targetPort: 19530
name: grpc
selector:
app: milvus
应用配置:
kubectl apply -f milvus-deployment.yaml
创建凭据 Secret:
kubectl create secret generic milvus-credentials \
--from-literal=password='Milvus' # Milvus 默认密码
对于生产部署,建议使用:
- Milvus Operator
- Milvus Helm Chart
- Zilliz Cloud(托管 Milvus 服务)
对于生产环境的缓存后端:
- 使用持久卷(而不是 emptyDir)
- 启用身份验证和 TLS
- 合理配置资源限制
- 设置监控和告警
- 使用 operator 或 Helm chart 简化管理
- 考虑使用托管服务以降低运维开销
Embedding 模型
语义缓存支持使用不同的 embedding 模型计算相似度:
- bert(默认):轻量级,384 维,适合通用场景
- qwen3:质量更高,1024 维,准确性更好
- gemma:较为均衡,768 维,性能适中
配置方式如下:
spec:
config:
global:
stores:
semantic_cache:
embedding_model: mmbert # 或通过显式覆盖 model_catalog 使用 qwen3、gemma
注意: 请确保缓存配置中的 dimension 与所选 embedding 模型一致。
后端迁移
从内存缓存迁移到 Redis 或 Milvus 的流程很简单:
- 在集群中部署 Redis 或 Milvus
- 创建凭据 Secret
- 使用新的后端配置更新 SemanticRouter CR
- 应用变更,operator 将执行滚动更新
迁移后缓存为空, 但会随着查询处理自然填充。
缓存配置参考
如需详细的配置选项,请使用 kubectl explain:
# Redis 缓存配置
kubectl explain semanticrouter.spec.config.global.stores.semantic_cache.redis
# Milvus 缓存配置
kubectl explain semanticrouter.spec.config.global.stores.semantic_cache.milvus
# HNSW 配置
kubectl explain semanticrouter.spec.config.global.stores.semantic_cache.hnsw
Semantic Router 配置
完整的 semantic router 配置嵌入在 CR 中。请参阅上方的完整示例以及 deploy/operator/config/samples/。
主要配置部分如下:
spec:
config:
providers:
defaults:
reasoning_families:
deepseek:
type: "chat_template_kwargs"
parameter: "thinking"
qwen3:
type: "chat_template_kwargs"
parameter: "enable_thinking"
gpt:
type: "reasoning_effort"
parameter: "reasoning_effort"
global:
model_catalog:
embeddings:
semantic:
mmbert_model_path: "models/mom-embedding-ultra"
use_cpu: true
embedding_config:
model_type: "mmbert"
preload_embeddings: true
target_dimension: 768
target_layer: 22
top_k: 1
min_score_threshold: 0.5
system:
prompt_guard: "models/mmbert32k-jailbreak-detector-merged"
domain_classifier: "models/mmbert32k-intent-classifier-merged"
pii_classifier: "models/mmbert32k-pii-detector-merged"
modules:
prompt_guard:
enabled: true
model_id: "models/mmbert32k-jailbreak-detector-merged"
threshold: 0.7
use_cpu: true
use_mmbert_32k: true
classifier:
domain:
model_id: "models/mmbert32k-intent-classifier-merged"
threshold: 0.5
use_cpu: true
use_mmbert_32k: true
pii:
model_id: "models/mmbert32k-pii-detector-merged"
threshold: 0.9
use_cpu: true
use_mmbert_32k: true
stores:
semantic_cache:
enabled: true
backend_type: "memory" # 或 redis、milvus、hybrid
similarity_threshold: 0.8
max_entries: 1000
ttl_seconds: 3600
eviction_policy: "fifo"
integrations:
tools:
enabled: true
top_k: 3
similarity_threshold: 0.2
tools_db_path: "config/tools_db.json"
fallback_to_empty: true
services:
api:
batch_classification:
max_batch_size: 100
concurrency_threshold: 5
max_concurrency: 8
metrics:
enabled: true
detailed_goroutine_tracking: true
sample_rate: 1.0
observability:
tracing:
enabled: false
provider: "opentelemetry"
exporter:
type: "otlp"
endpoint: "jaeger:4317"
工具数据库
定义可供自动选择的工具:
spec:
toolsDb:
- tool:
type: "function"
function:
name: "search_web"
description: "Search the web for information"
parameters:
type: "object"
properties:
query:
type: "string"
description: "Search query"
required: ["query"]
description: "Search the internet, web search, find information online"
category: "search"
tags: ["search", "web", "internet"]
- tool:
type: "function"
function:
name: "calculate"
description: "Perform mathematical calculations"
parameters:
type: "object"
properties:
expression:
type: "string"
required: ["expression"]
description: "Calculate mathematical expressions"
category: "math"
tags: ["math", "calculation"]
自动扩缩容(HPA)
spec:
autoscaling:
enabled: true
minReplicas: 2
maxReplicas: 10
targetCPUUtilizationPercentage: 70
targetMemoryUtilizationPercentage: 80
Ingress 配置
spec:
ingress:
enabled: true
className: "nginx" # 或 "haproxy"、"traefik" 等
annotations:
cert-manager.io/cluster-issuer: "letsencrypt-prod"
hosts:
- host: router.example.com
paths:
- path: /
pathType: Prefix
servicePort: 8080
tls:
- secretName: router-tls
hosts:
- router.example.com
生产部署
高可用设置
apiVersion: vllm.ai/v1alpha1
kind: SemanticRouter
metadata:
name: prod-router
spec:
replicas: 3
# 用反亲和性将 pod 分散到不同节点
affinity:
podAntiAffinity:
requiredDuringSchedulingIgnoredDuringExecution:
- labelSelector:
matchLabels:
app.kubernetes.io/instance: prod-router
topologyKey: kubernetes.io/hostname
# 自动扩缩容
autoscaling:
enabled: true
minReplicas: 3
maxReplicas: 20
targetCPUUtilizationPercentage: 70
# 生产环境资源
resources:
limits:
memory: "10Gi"
cpu: "4"
requests:
memory: "5Gi"
cpu: "2"
# 严格的探针配置
livenessProbe:
enabled: true
initialDelaySeconds: 60
periodSeconds: 30
failureThreshold: 3
readinessProbe:
enabled: true
initialDelaySeconds: 30
periodSeconds: 10
failureThreshold: 3
Pod Disruption Budget
创建 PDB,确保更新期间的可用性:
apiVersion: policy/v1
kind: PodDisruptionBudget
metadata:
name: prod-router-pdb
spec:
maxUnavailable: 1
selector:
matchLabels:
app.kubernetes.io/instance: prod-router
资源分配指南
| 工作负载类型 | 内存请求 | CPU 请求 | 内存限制 | CPU 限制 |
|---|---|---|---|---|
| 开发 | 1Gi | 500m | 2Gi | 1 |
| 预发布 | 3Gi | 1 | 7Gi | 2 |
| 生产 | 5Gi | 2 | 10Gi | 4 |
监控与可观测性
指标
Prometheus 指标通过 9190 端口暴露:
# 通过端口转发在本地访问指标
kubectl port-forward svc/my-router 9190:9190
# 查看指标
curl http://localhost:9190/metrics
关键指标:
| 指标族 | 示例指标 |
|---|---|
| 请求 | llm_model_requests_total, llm_request_errors_total |
| 错误 | llm_request_errors_total{reason="timeout"} |
| 延迟 | llm_model_completion_latency_seconds, llm_model_ttft_seconds, llm_model_tpot_seconds, llm_model_routing_latency_seconds |
| Token 与成本 | llm_model_tokens_total, llm_model_prompt_tokens_total, llm_model_completion_tokens_total, llm_model_cost_total |
| 路由 | llm_model_routing_modifications_total, llm_routing_reason_codes_total |
| 选择 | llm_model_selection_total, llm_model_selection_duration_seconds, llm_model_inflight_requests |
| 缓存 | llm_cache_plugin_hits_total, llm_cache_plugin_misses_total, llm_cache_warmth_estimate |
| RAG | rag_retrieval_attempts_total, rag_retrieval_latency_seconds, rag_cache_hits_total, rag_cache_misses_total |
| 会话 | llm_session_model_transitions_total, llm_session_turn_prompt_tokens, llm_session_turn_completion_tokens, llm_session_turn_cost |
| 翻译与请求参数策略 | llm_translation_lossy_total, sr_request_params_blocked_total, sr_request_params_unknown_field_stripped_total |
ServiceMonitor(Prometheus Operator)
apiVersion: monitoring.coreos.com/v1
kind: ServiceMonitor
metadata:
name: semantic-router-metrics
spec:
selector:
matchLabels:
app.kubernetes.io/instance: my-router
endpoints:
- port: metrics
interval: 30s
path: /metrics
分布式追踪
启用 OpenTelemetry tracing:
spec:
config:
observability:
tracing:
enabled: true
provider: "opentelemetry"
exporter:
type: "otlp"
endpoint: "jaeger-collector:4317"
insecure: true
sampling:
type: "probabilistic"
rate: 0.1
推荐使用 probabilistic 采样类型。为兼容现有配置,也接 受 traceidratio 和 trace_id_ratio 作为别名。
故障排除
常见问题
后端发现失败
症状: 日志中出现 "No backends found" 或 "Failed to discover backend"
对于 KServe 后端:
# 检查 InferenceService 是否存在并已就绪
kubectl get inferenceservice llama-3-8b
# 检查 KServe 是否已创建 predictor service
kubectl get service llama-3-8b-predictor
# 验证 InferenceService 状态
kubectl describe inferenceservice llama-3-8b
对于 Llama Stack 后端:
# 验证具有正确 label 的 service 是否存在
kubectl get services -l app=llama-stack,model=llama-3.3-70b
# 检查 service label 是否与 CR 中的 discoveryLabels 一致
kubectl get service <service-name> -o jsonpath='{.metadata.labels}'
对于直连 service 后端:
# 验证 service 是否存在于指定 namespace
kubectl get service vllm-deepseek -n vllm-serving
# 检查 service 是否定义了端口
kubectl get service vllm-deepseek -n vllm-serving -o jsonpath='{.spec.ports[0]}'
Gateway 集成问题
症状: 尚不存在 HTTPRoute,或流量未到达 semantic router
当前 operator 实现不会自动创建 HTTPRoute 资源。如果你尚未自行应用 HTTPRoute manifest,这属于预期行为,并非 reconcile 失败。
# 验证 Gateway 是否存在
kubectl get gateway istio-ingressgateway -n istio-system
# 检查你是否自行创建了 HTTPRoute
kubectl get httproute -l app.kubernetes.io/instance=my-router
# 验证 Gateway 是否支持 HTTPRoute(Gateway API v1)
kubectl get gateway istio-ingressgateway -n istio-system -o yaml | grep -A5 listeners
# 检查 operator 状态
kubectl get semanticrouter my-router -o jsonpath='{.status.gatewayMode}'
# 应显示:"gateway-integration"
# 检查 operator 日志中的当前占位消息
kubectl logs deployment/semantic-router-operator-controller-manager -n semantic-router-system | grep 'HTTPRoute creation placeholder'
如果 operator 日志显示 HTTPRoute creation placeholder - requires Gateway API version-specific implementation,说明 controller 已识别 gateway mode,但没有创建 route。请手动应用 HTTPRoute manifest,并验证它是否以 8080 端口上的 semantic router service 为目标。
OpenShift Route 问题
症状: OpenShift 上未创建 Route
# 验证是否运行在 OpenShift 集群上
kubectl api-resources | grep route.openshift.io
# 检查 Route 是否已创建
kubectl get route -l app.kubernetes.io/instance=my-router
# 检查 operator 是否探测到 OpenShift
kubectl logs -n semantic-router-operator-system \
deployment/semantic-router-operator-controller-manager \
| grep -i "openshift\|route"
# 验证 Route 状态
kubectl get semanticrouter my-router -o jsonpath='{.status.openshiftFeatures}'
Pod 卡在 ImagePullBackOff
# 检查镜像拉取 Secret
kubectl describe pod <pod-name>
# 创建镜像拉取 Secret
kubectl create secret docker-registry ghcr-secret \
--docker-server=ghcr.io \
--docker-username=<username> \
--docker-password=<personal-access-token>
# 添加到 SemanticRouter
spec:
imagePullSecrets:
- name: ghcr-secret
PVC 卡在 Pending
# 检查 StorageClass 是否存在
kubectl get storageclass
# 检查 PVC event
kubectl describe pvc my-router-models
# 更新 CR 中的 StorageClass
spec:
persistence:
storageClassName: "your-available-storage-class"
模型无法下载
# 检查 HF token Secret 是否存在
kubectl get secret hf-token-secret
# 创建 HF token Secret
kubectl create secret generic hf-token-secret \
--from-literal=token=hf_xxxxxxxxxxxxx
# 添加到 SemanticRouter CR
spec:
env:
- name: HF_TOKEN
valueFrom:
secretKeyRef:
name: hf-token-secret
key: token
operator 未正确探测平台
# 检查 operator 日志中的平台探测信息
kubectl logs -n semantic-router-operator-system \
deployment/semantic-router-operator-controller-manager \
| grep -i "platform\|openshift"
# 应看到以下消息之一:
# "Detected OpenShift platform - will use OpenShift-compatible security contexts"
# "Detected standard Kubernetes platform - will use standard security contexts"
开发
本地开发
cd deploy/operator
# 运行测试
make test
# 生成 CRD 和代码
make generate
make manifests
# 构建 operator 二进制文件
make build
# 使用你的 kubeconfig 在本地运行
make run
使用 kind 测试
# 创建 kind 集群
kind create cluster --name operator-test
# 构建并加载镜像
make docker-build IMG=semantic-router-operator:dev
kind load docker-image semantic-router-operator:dev --name operator-test
# 部署
make install
make deploy IMG=semantic-router-operator:dev
# 创建测试实例
kubectl apply -f config/samples/vllm_v1alpha1_semanticrouter.yaml