---

name: deploy-ml-model-serving locale: wenyan-ultra source_locale: en source_commit: 82c77053 fence_basis_commit: 82c77053 translator: "Julius Brussee homage — caveman" translation_date: "2026-04-19" description: > Deploy machine learning models to production serving infrastructure using MLflow, BentoML, or Seldon Core with REST/gRPC endpoints, implement autoscaling, monitoring, and A/B testing capabilities for high-performance model inference at scale. Use when deploying trained models for real-time inference, setting up REST or gRPC prediction APIs, implementing autoscaling for variable load, running A/B tests between model versions, or migrating from batch to real-time inference. license: MIT allowed-tools: Read Write Edit Bash Grep Glob metadata: author: Philipp Thoss version: "1.0" domain: mlops complexity: advanced language: multi tags: model-serving, bentoml, seldon, rest-api, grpc

部署 ML 模型服務

詳例見 Extended Examples。

部 ML 模型於生產：可擴服務設施、監控、A/B 測試。

用

• 訓好模型→生產實時推理
• REST/gRPC 預測 API
• 變載自動擴
• 模型版本 A/B
• 批→實時遷移
• 構低時延預測服務
• 理多版本生產

入

• 必：MLflow 註冊模型或訓好物
• 必：Kubernetes 集群或容編平
• 必：服務框架（MLflow、BentoML、Seldon Core、TorchServe）
• 可：GPU（深度學習）
• 可：監設施（Prometheus、Grafana）
• 可：負載均衡+入控

法

一：以 MLflow 部署

速部 scikit-learn、PyTorch、TensorFlow。

# Serve model locally for testing
mlflow models serve \
  --model-uri models:/customer-churn-classifier/Production \
  --port 5001 \
  --host 0.0.0.0

# Test endpoint
curl -X POST http://localhost:5001/invocations \
  -H 'Content-Type: application/json' \
  -d '{
    "dataframe_records": [
      {"feature1": 1.0, "feature2": 2.0, "feature3": 3.0}
    ]
  }'

Docker 部署：

# Dockerfile.mlflow-serving
FROM python:3.9-slim

# Install MLflow and dependencies
RUN pip install mlflow boto3 scikit-learn

# Set environment variables
ENV MLFLOW_TRACKING_URI=http://mlflow-server:5000
# ... (see EXAMPLES.md for complete implementation)

Docker Compose 本地測試：

# docker-compose.mlflow-serving.yml
version: '3.8'

services:
  model-server:
    build:
      context: .
      dockerfile: Dockerfile.mlflow-serving
# ... (see EXAMPLES.md for complete implementation)

測部署：

# test_mlflow_serving.py
import requests
import json

def test_prediction():
    url = "http://localhost:8080/invocations"

    # Prepare input data
# ... (see EXAMPLES.md for complete implementation)

得： 模型服務啟，應 HTTP POST，以 JSON 返預測，Docker 容器無錯。

敗： 查模型 URI（mlflow models list），驗 MLflow 跟蹤服務可達，確容器內依賴齊，查端口（netstat -tulpn | grep 8080），驗模型風味相容，察容器日誌（docker logs <container-id>）。

二：以 BentoML 部署生產

BentoML 高級服務。

# bentoml_service.py
import bentoml
from bentoml.io import JSON, NumpyNdarray
import numpy as np
import pandas as pd

# Load model from MLflow
import mlflow
# ... (see EXAMPLES.md for complete implementation)

構+容器化：

# Build Bento
bentoml build

# Containerize
bentoml containerize customer_churn_classifier:latest \
  --image-tag customer-churn:v1.0

# Run container
docker run -p 3000:3000 customer-churn:v1.0

BentoML 配置：

# bentofile.yaml
service: "bentoml_service:ChurnPredictionService"
include:
  - "bentoml_service.py"
  - "preprocessing.py"
python:
  packages:
    - scikit-learn==1.0.2
    - pandas==1.4.0
    - numpy==1.22.0
    - mlflow==2.0.1
docker:
  distro: debian
  python_version: "3.9"
  cuda_version: null  # Set to "11.6" for GPU support

Kubernetes 部署：

# k8s/deployment.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
  name: churn-prediction
  labels:
    app: churn-prediction
spec:
# ... (see EXAMPLES.md for complete implementation)

部至 Kubernetes：

# Apply Kubernetes manifests
kubectl apply -f k8s/deployment.yaml

# Check deployment status
kubectl get deployments
kubectl get pods
kubectl get services

# Test endpoint
EXTERNAL_IP=$(kubectl get svc churn-prediction-service -o jsonpath='{.status.loadBalancer.ingress[0].ip}')
curl -X POST http://$EXTERNAL_IP/predict \
  -H 'Content-Type: application/json' \
  -d '{"instances": [{"tenure": 12, "monthly_charges": 70.35}]}'

得： BentoML 服務構成，容器跑並服預測，K8s 部署建 3 副本，負載均衡露外，健康檢通。

敗： 驗 BentoML 裝（bentoml --version），查模型於 BentoML 庫（bentoml models list），確 Docker 守護跑，驗 K8s 集群訪（kubectl cluster-info），查資源限，察 pod 日誌（kubectl logs <pod-name>），驗選擇器匹 pod 標籤。

三：以 Seldon Core 行進階功能

多模服務、A/B、可解釋。

# seldon_wrapper.py
import logging
from typing import Dict, List, Union
import numpy as np
import mlflow

logger = logging.getLogger(__name__)

# ... (see EXAMPLES.md for complete implementation)

Seldon 部署配：

# seldon-deployment.yaml
apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: churn-classifier
  namespace: seldon
spec:
  name: churn-classifier
# ... (see EXAMPLES.md for complete implementation)

A/B 測試配：

# seldon-ab-test.yaml
apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: churn-classifier-ab
spec:
  name: churn-classifier-ab
  predictors:
# ... (see EXAMPLES.md for complete implementation)

部至 Kubernetes：

# Install Seldon Core operator
kubectl create namespace seldon-system
helm install seldon-core seldon-core-operator \
  --repo https://storage.googleapis.com/seldon-charts \
  --namespace seldon-system \
  --set usageMetrics.enabled=true

# Create namespace for models
# ... (see EXAMPLES.md for complete implementation)

得： Seldon 算子裝成，模部署建 pod，REST 應預測，A/B 流量分對，Seldon 分析錄指標。

敗： 驗 Seldon 算子跑（kubectl get pods -n seldon-system），查 SeldonDeployment 狀態（kubectl describe seldondeployment），確鏡像倉可達，驗模型 URI 解析，查 RBAC 權限，察模容器日誌。

四：行監與可觀

加監控於模型服務設施。

# monitoring.py
from prometheus_client import Counter, Histogram, Gauge, start_http_server
import time
import logging

logger = logging.getLogger(__name__)

# Prometheus metrics
# ... (see EXAMPLES.md for complete implementation)

Prometheus 配：

# prometheus-config.yaml
global:
  scrape_interval: 15s
  evaluation_interval: 15s

scrape_configs:
  - job_name: 'model-serving'
    kubernetes_sd_configs:
# ... (see EXAMPLES.md for complete implementation)

Grafana 儀板 JSON：

{
  "dashboard": {
    "title": "ML Model Serving Metrics",
    "panels": [
      {
        "title": "Predictions Per Second",
        "targets": [
          {
# ... (see EXAMPLES.md for complete implementation)

得： Prometheus 抓成，Grafana 儀板示實時吞吐、時延百分位、錯率、活躍請求。

敗： 驗 Prometheus 抓目標 UP（http://prometheus:9090/targets），查指標端點可達（curl http://model-pod:8000/metrics），確 K8s 服務發現配，驗 Grafana 數據源連，查防火牆指標端口。

五：行自動擴

配水平 pod 自動擴於請求負載。

# hpa.yaml
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: churn-prediction-hpa
  namespace: seldon
spec:
  scaleTargetRef:
# ... (see EXAMPLES.md for complete implementation)

施自動擴：

# Enable metrics server (if not already installed)
kubectl apply -f https://github.com/kubernetes-sigs/metrics-server/releases/latest/download/components.yaml

# Apply HPA
kubectl apply -f hpa.yaml

# Check HPA status
kubectl get hpa -n seldon
kubectl describe hpa churn-prediction-hpa -n seldon

# Load test to trigger scaling
kubectl run -it --rm load-generator --image=busybox --restart=Never -- /bin/sh -c "while sleep 0.01; do wget -q -O- http://churn-prediction-service/predict; done"

# Watch scaling
kubectl get hpa -n seldon --watch

得： HPA 監 CPU/內存/自定指標，負載擴，穩定後縮，min/max 守。

敗： 驗 metrics-server 跑（kubectl get deployment metrics-server -n kube-system），查 pod 資源請求定（HPA 需），確自定指標可用（若用），驗 HPA 控制器 RBAC，查穩定窗不過嚴。

六：行金絲雀部署

漸進新模版本。

# canary-deployment.yaml
apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
  name: churn-classifier-canary
spec:
  name: churn-classifier-canary
  predictors:
# ... (see EXAMPLES.md for complete implementation)

漸進推出腳本：

# canary_rollout.py
import time
import subprocess
import logging

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# ... (see EXAMPLES.md for complete implementation)

得： 金絲雀始 0% 流量，自動漸進轉移，各階健康檢通，指標劣即回滾，諸階通畢畢。

敗： 驗 Seldon 有多預測器，查流量比和 100，確金絲雀像存可拉，驗 Prometheus 指標可用於健康檢，查回滾邏輯執，察兩版 pod 日誌。

驗

• 模型服務應預測
• REST/gRPC 可用並錄文
• Docker 容器構跑
• K8s 部署建期望副本
• 負載均衡露外
• 健康檢（活躍/就緒）通
• Prometheus 指標出+抓
• Grafana 示實時
• 自動擴觸
• A/B 流量分對
• 金絲雀漸進推出
• 金絲雀失時回滾工作

忌

• 冷啟時延：首請慢因載模→用就緒探+適延，行模緩存
• 內存洩漏：長跑累→監內存，週期重啟，剖碼
• 依賴衝突：與服務框架不容→精釘，Docker 前測
• 資源限過低：OOM/節流→剖用量，按壓測設限
• 缺健康檢：K8s 路不健康 pod→行活躍/就緒探
• 無回滾策：壞部無易回→用金絲雀，留前版
• 忽時延：只顧準非速→測時延，優化模/碼，批
• 單副本：無高可用，部署時宕→至少 2 副本，反親和
• 無監：至客訴始知→首日即監
• GPU 未用：有而不用→設 CUDA 可見，驗 K8s GPU 分配

參

• register-ml-model
• run-ab-test-models
• deploy-to-kubernetes
• monitor-ml-model-performance
• orchestrate-ml-pipeline