Model Deployment

Model Deployment

Purpose

Deploy trained ML models as production services with proper API design, containerization, and scaling.

How It Works

Step 1: Package Model

• Serialize model (joblib, pickle, ONNX, TorchScript)
• Define input/output schema (Pydantic models)
• Create prediction function with preprocessing

Step 2: Build API

| Framework | Best For | |-----------|----------| | FastAPI | Modern, async, auto-docs | | Flask | Simple, widely known | | BentoML | ML-specific, batching | | Ray Serve | Scalable, composable |

Step 3: Containerize

• Dockerfile with minimal base image
• Multi-stage build for smaller images
• Health check endpoint
• Environment variable configuration

Step 4: Deploy

| Target | Complexity | Scaling | |--------|-----------|---------| | Docker Compose | Low | Manual | | AWS SageMaker | Medium | Auto | | GCP Vertex AI | Medium | Auto | | Kubernetes | High | Full control | | Serverless (Lambda) | Low | Auto |

Step 5: Test & Monitor

• Load testing (locust, k6)
• Latency and throughput benchmarks
• Error handling and fallbacks

Usage Examples

"Deploy my scikit-learn model as a FastAPI endpoint in Docker"

"Set up a SageMaker endpoint for real-time inference"

Output Format

• API Code: FastAPI/Flask application
• Dockerfile: Production-ready container
• Deployment Config: Cloud/K8s configuration
• Test Script: API testing and load testing code

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ericwang915
• Source: ericwang915/data-scientist-skills
• License: MIT

Install and usage instructions live in the source repository linked above.