Deploy to Azure Function App - serverless approach

Following this tutorial.

Steps:

1. Setup
2. Deploy model
3. Test model
4. Delete endpoint

Model to be deployed:

model_path = "../../models/model.pkl"

resource_group = "azure-cognitive-services-accelerator"

1. Setup

1.1 Setup Azure Functions CLI

Install Azure Functions CLI package following the docs (you may need xcode-select --install too):

!brew tap azure/functions

!brew install azure-functions-core-tools@4

1.2 Setup location Azure Function

Initialise local function project folder:

!func init sample_project --python

%cd sample_project

Create new function that is triggered by HTTP:

!func new --name sample_endpoint --template "HTTP trigger" --authlevel "anonymous" --worker-runtime python

Copy scikit-learn model and add inferencing script + dependencies. Make sure scikit-learn is same version in which model was created otherwise may get compatibility errors (e.g. see here for common scikit-learn error and see here for a python3.8 vs 3.9 issue for older sklearn versions)

!cp $model_path sample_endpoint

%%writefile -a requirements.txt

scikit-learn==1.2.0

%%writefile sample_endpoint/__init__.py
import azure.functions as func
import numpy as np
import logging, sklearn, joblib, json

def main(req: func.HttpRequest) -> func.HttpResponse:
    logging.info(sklearn.__version__) # check this matches sklearn version in which model was trained
    pred = joblib.load('sample_endpoint/model.pkl').predict(np.array(req.get_json()['data']))
    return func.HttpResponse(json.dumps({"prediction": pred.tolist()}))

Optional: test model locally using func start in a new terminal and query local API with requests in Test section below.

1.3 Create resources: Azure Function App and Storage Account

Create storage account under appropriate resource group (ideally keep all assets under same resource group for cost tracking).

Note that storage account and function app can’t include _ in their names, and function app creation fails silently if you don’t follow this.

!az storage account create --name sampleprojectstorage --resource-group $resource_group --sku Standard_LRS

Create function app under same resource group and location. Check that runtime-version is the same as python version in which model was created/compatible with the sklearn version.

!az functionapp create --resource-group $resource_group --consumption-plan-location uksouth --runtime python --runtime-version 3.10 --functions-version 4 --name sampleprojectapp --os-type linux --storage-account sampleprojectstorage

Use !az functionapp update --resource-group $resource_group --name sample_project to rerun the above command.

2. Deploy model

Publish function app (i.e. deploy). If first time results in Deploy Failed, just try rerunning the command (shrugs shoulders). Take note of the invoke URL at the end for API queries.

!func azure functionapp publish sampleprojectapp

endpoint_url = "https://sampleprojectapp.azurewebsites.net/api/sample_endpoint"

3. Test endpoint

You can manage and monitor the logs of the function app on the Azure portal under Function App.

import requests
import json
import numpy as np

input_payload = json.dumps({
    'data': np.load("../../data/diabetes.npz")["X_test"][:2].tolist(),
})

requests.post(endpoint_url, input_payload, headers={'Content-Type':'application/json'}).json()

4. Delete endpoint

Delete created resources. Or you could do this in Azure portal.

!az functionapp delete --name sampleprojectapp --resource-group $resource_group

!az storage account delete --name sampleprojectstorage --resource-group $resource_group -y

Optional: delete local files

%cd ..

!rm -r sample_project