model-deployment（模型部署）

model-deployment

Purpose

This skill automates the deployment of machine learning models to production environments using containers (e.g., Docker) and orchestration tools (e.g., Kubernetes), ensuring scalable and reliable ML model serving.

When to Use

• When you need to containerize and deploy a trained ML model for real-time inference in production.
• For updating existing deployments in response to model retraining or performance issues.
• In MLOps pipelines where models must be versioned, monitored, and rolled back easily.
• When integrating with cloud providers like AWS EKS or Google GKE for managed orchestration.

Key Capabilities

• Builds Docker images from model artifacts and deploys them to Kubernetes clusters.
• Supports model versioning via tags and handles rolling updates for zero-downtime deployments.
• Integrates with ML frameworks like TensorFlow or PyTorch for serving models via APIs.
• Manages resource allocation, such as CPU/GPU requests in Kubernetes pods, e.g., resources: limits: cpu: 2.
• Automates scaling based on traffic, using Kubernetes Horizontal Pod Autoscalers.

Usage Patterns

To use this skill, first prepare your model in a Docker-friendly format, then build and deploy it. Always set environment variables for authentication, like $KUBECONFIG for Kubernetes access.

Pattern 1: Basic deployment

• Export your model as a saved file (e.g., model.h5) and write a Dockerfile.
• Build the image locally or in CI/CD.
• Apply a Kubernetes deployment YAML to orchestrate the container.

Pattern 2: Update an existing deployment

• Tag a new model version and rebuild the Docker image.
• Use kubectl to apply changes, specifying the new image tag.
• Monitor the rollout and roll back if needed using built-in commands.

Always verify cluster access before starting; check with kubectl get nodes to ensure connectivity.

Common Commands/API

Use these CLI commands for core operations. For API interactions, reference Kubernetes REST API endpoints.

• Build and tag a Docker image: docker build -t mymlmodel:v1. This creates an image from the current directory.
• Push the image to a registry: docker push mymlmodel:v1 Requires authentication via $DOCKER_REGISTRY_TOKEN as an env var.
• Deploy to Kubernetes: kubectl apply -f deployment.yaml Where deployment.yaml includes: apiVersion: apps/v1 kind: Deployment metadata: name: myml-deployment spec: replicas: 3, template: spec: containers: - name: model-server image: mymlmodel:v1
• Scale the deployment: kubectl scale deployment myml-deployment --replicas=5
• API endpoint for querying deployments: Use the Kubernetes API at GET /apis/apps/v1/namespaces/default/deployments with authentication via bearer token in $KUBE_API_TOKEN.

For config formats, use Kubernetes YAML files, e.g.:

apiVersion: v1
kind: Service
metadata: name: model-service
spec: selector: app: mymlmodel, ports: - protocol: TCP port: 80

Integration Notes

Integrate this skill with CI/CD tools like GitHub Actions or Jenkins by adding steps in your pipeline YAML. For example, in GitHub Actions:

jobs:
  deploy:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v2
      - run: docker build -t mymlmodel:${{ github.sha }} .
      - run: kubectl apply -f k8s/deployment.yaml --context=$KUBE_CONTEXT

Set env vars for secrets, e.g., $GITHUB_TOKEN for repo access and $KUBE_CONTEXT for cluster selection. Ensure your ML pipeline outputs are in a standard format, like a pickled model file, for seamless Docker integration.

Error Handling

Handle common errors proactively. If docker build fails with "no such file," verify the Dockerfile path and required files. For Kubernetes errors like "image pull failed," check image registry credentials via $DOCKER_REGISTRY_TOKEN.

• Error: Pod not ready – Fix by inspecting logs with kubectl logs <pod-name> and ensure resources match in deployment YAML, e.g., add resources: requests: memory: "1Gi".
• Error: Authentication failure – Set env vars correctly, e.g., export KUBECONFIG=~/.kube/config, and test with kubectl get pods.
• For API errors, like 401 Unauthorized, retry with refreshed tokens from $KUBE_API_TOKEN and use exponential backoff in scripts.

Always include try-catch in automation scripts, e.g.:

import subprocess
try:
    subprocess.run(["kubectl", "apply", "-f", "deployment.yaml"], check=True)
except subprocess.CalledProcessError as e:
    print(f"Deployment failed: {e}")

Graph Relationships

• Related Cluster: aimlops
• Related Tags: mlops, deployment, containers
• Connected Skills: model-training (for pre-deployment), monitoring (for post-deployment observability)