Skip to content

AZURE_DEPLOYMENT.md

Latest commit

 

History

History
100 lines (76 loc) · 2.41 KB

AZURE_DEPLOYMENT.md

File metadata and controls

100 lines (76 loc) · 2.41 KB

Azure AutoML Real-Time Deployment

Microsoft's recommended approach for deploying AutoML models to production for real-time endpoints is to download your AutoML model and deploy it using Azure Functions.

Prerequisites

  • Azure CLI installed and logged in
  • Azure Functions Core Tools installed
  • Python 3.8+ installed locally
  • Access to the Azure ML workspace containing your model
# Install Azure Functions Core Tools (Mac)
brew tap azure/functions
brew install azure-functions-core-tools@4

# Verify installation
func --version

Deployment

Step 1: Download Your AutoML Model

# Download your trained AutoML model
az ml model download \
  --name titanic-survival \
  --version 1 \
  --resource-group ubc-cdl10 \
  --workspace-name UBC-CDL10 \
  --download-path ./titanic-survival-app/automl-model/

Step 2: Create Azure Function App

# Create a new storage account (name must be globally unique, 3-24 chars, lowercase/numbers only)
az storage account create \
  --name ubccdl10titanic \
  --resource-group ubc-cdl10 \
  --location eastus2 \
  --sku Standard_LRS

# Create Function App
az functionapp create \
  --resource-group ubc-cdl10 \
  --consumption-plan-location eastus2 \
  --runtime python \
  --runtime-version 3.8 \
  --functions-version 4 \
  --name titanic-survival-app \
  --storage-account ubccdl10titanic

Step 3: Create Function Code

Step 4: Deploy Function

cd titanic-survival-app
func azure functionapp publish titanic-survival-app

Step 5: Test Your Deployment

# Test the function
curl -X POST https://titanic-survival-app.azurewebsites.net/api/predict \
  -H "Content-Type: application/json" \
  -d '{"data": [[3, 1, 22, 1, 0, 7.25, 2]]}'

Expected response:

{
  "prediction": [0]
}

🔧 Troubleshooting

Check Function Logs

# Check function logs
az functionapp log tail --name titanic-survival-app --resource-group ubc-cdl10

Common Issues

  1. Model Loading Errors: Ensure model.pkl is in the same directory as init.py
  2. Dependency Issues: Pin all package versions in requirements.txt
  3. Memory Issues: Use Premium plan for large models
  4. JSON Issues: Ensure input data format matches training data