Model Deployment Quick Reference Guide

Quick reference for deploying ML models.

Table of Contents

• Model Serialization
• API Frameworks
• Docker Commands
• Cloud Platforms
• Common Issues & Solutions
• Best Practices Checklist

Model Serialization

Save Models

# Scikit-learn
joblib.dump(model, 'model.joblib')

# Keras
model.save('model.h5')

# PyTorch
torch.save(model.state_dict(), 'model.pth')

API Frameworks

FastAPI

from fastapi import FastAPI
app = FastAPI()

@app.post("/predict")
async def predict(features: list[float]):
    return {"prediction": model.predict([features])[0]}

Flask

from flask import Flask, jsonify
app = Flask(__name__)

@app.route('/predict', methods=['POST'])
def predict():
    return jsonify({'prediction': model.predict([features])[0]})

Docker Commands

# Build
docker build -t ml-api .

# Run
docker run -p 8000:8000 ml-api

# Compose
docker-compose up

A/B Testing

Statistical Significance

from scipy import stats
import numpy as np

def analyze_ab_test(conversions_a, total_a, conversions_b, total_b):
    """Z-test for two proportions"""
    rate_a = conversions_a / total_a
    rate_b = conversions_b / total_b
    
    # Pooled proportion
    pooled = (conversions_a + conversions_b) / (total_a + total_b)
    se = np.sqrt(pooled * (1 - pooled) * (1/total_a + 1/total_b))
    
    # Z-score and p-value
    z = (rate_b - rate_a) / se
    p_value = 2 * (1 - stats.norm.cdf(abs(z)))
    
    return {
        'rate_a': rate_a,
        'rate_b': rate_b,
        'p_value': p_value,
        'significant': p_value < 0.05
    }

Best Practices

• Clear hypothesis: Define what you're testing
• Primary metric: One clear success metric
• Sample size: Ensure statistical power
• Don't peek early: Avoid stopping prematurely
• Monitor: Track deployed model performance

Cloud Platforms

Common Issues & Solutions

Issue 1: Model Too Large

Solution: Quantize, prune, use smaller models

Issue 2: Slow Response

Solution: Optimize model, use caching, batch processing

Best Practices Checklist

• Validate all inputs
• Handle errors gracefully
• Add health checks
• Use HTTPS in production
• Monitor performance
• Version your models
• Document API

Try next: Hit your local /predict once, then add a /health check and one latency log line.