Part 1

.gitignore

@@ -12,4 +12,7 @@ node_modules/
 .serverless/
 
 # Env files
-*.env
+*.env
+
+# Python
+venv/

jobs/ml-ops/README.md

@@ -1,67 +1,96 @@
 # Serverless MLOps
 
-In this example, we train and deploy a binary classification inference API using serverless computing resources (job+container). We use object storage resources to store data and training artifacts. We use container registry to store docker images.
+In this example, we train and deploy a binary classification inference model using Scaleway Serverless. To do this, we use the following resources:
 
-## Use case: Bank Telemarketing
+1. Serverless Job for training
+2. Serverless Job to populate data in S3
+3. Serverless Container for inference
 
-### Context
+We use object storage to share data between the two.
+
+## Context
+
+In this example we use a bank telemarketing dataset to predict if a client would engage in a term deposit subscription.
+
+This dataset records marketing phone calls made to clients. The outcome of the phone call is in shown in the `y` column:
 
-We use a bank telemarketing dataset to predict if a client would engage in a term deposit subscription. This dataset records marketing phone calls made to clients. The outcome of the phone call is in shown in the `y` column:
 * `0` : no subscription
 * `1` : subscription
 
-### Data Source
+## Data Source
 
 The dataset has many versions and is open-sourced and published [here](http://archive.ics.uci.edu/dataset/222/bank+marketing) on the UCI Machine Leaning repository and is close to the one analyzed in the following research work:
 
 * [Moro et al., 2014] S. Moro, P. Cortez and P. Rita. A Data-Driven Approach to Predict the Success of Bank Telemarketing. Decision Support Systems, Elsevier, 62:22-31, June 2014
 
-We use the dataset labelled in the source as `bank-additional-full.csv`. You can download, extract this file, rename it to `bank_telemarketing.csv` then put it under this [directory](./s3/data-store/data/).
+## Running the example
 
-## How to deploy your MLOps pipeline on Scaleway Cloud?
-
-### Step A: Create cloud resources for the ML pipeline
+### Step 1. Provision resources with Terraform
 
-Create `.env` file in `jobs/data-loader-job` and `jobs/ml-job` directories and fill them as it follows:
+Set your Scaleway access key, secret key and project ID in environment variables:
 
-```text
-SCW_ACCESS_KEY=<access-key>
-SCW_SECRET_KEY=<secret-key>
+```console
+export TF_VAR_access_key=<your-access-key>
+export TF_VAR_secret_key=<your-secret-key>
+export TF_VAR_project_id=<your-project-id>
+
+cd terraform
+terraform init
+terraform plan
+terraform apply
 ```
 
-Create `.tfvars` file in `/terraform` directory and put variable values in it:
+### Step 2. Run the data and training Jobs
+
+*At the time of writing, the Scaleway CLI does not support Jobs, so we use a Python script*
 
 ```
-region = "fr-par"
-access_key = "<access-key>"
-secret_key = "<secret_key>"
-project_id = "<project_id>"
-data_file     = "bank_telemarketing.csv"
-model_object  = "classifier.pkl"
-image_version = "v1"
+cd scripts
+
+python3 -m venv venv
+source venv/bin/activate
+pip install -r requirements.txt
+
+python3 run upload
+python3 run training
 ```
 
-Then perform:
+You can then check your Job runs in the [Jobs Console](https://console.scaleway.com/serverless-jobs/jobs).
+
+### Step 4. Use the inference API
 
-```bash
-cd terraform
-terraform init
-terraform plan -var-file=testing.tfvars
-terraform apply -var-file=testing.tfvars
 ```
+export INFERENCE_URL=$(terraform output endpoint)
 
-### Step B: Define and run a job to ship data from public source to s3
+curl -X POST \
+  -H "Content-Type: application/json" \
+  -d @inference/example.json
+  $INFERENCE_URL
+```
 
-Use the console to define and run the data loader job using image pushed to Scaleway registry.
+## Local testing
 
-cf. this [readme](./jobs/data-loader-job/README.md)
+To test the example locally you can use [Docker Compose](https://docs.docker.com/compose/install/).
 
-### Step C: Define and run the ML job to train classifier
+```
+# Build the containers locally
+docker compose build
 
-Use the console to define and the ML job using image pushed to Scaleway registry.
+# Run the data job
+docker compose run data
 
-cf. this [readme](./jobs/ml-job/README.md)
+# Run the training
+docker compose run training
 
-### Step D: Call your serverless container to (re)load model and to get inference results 
+# Start the inference server
+docker compose up inference
+```
+
+Access the inference API locally:
 
-cf. this [readme](./containers/inference-api/README.md)
+```
+curl -X POST \
+  -H "Content-Type: application/json" \
+  -d @inference/example.json
+  http://localhost:8080
+```

jobs/ml-ops/data/.gitignore

@@ -0,0 +1 @@
+dataset/

jobs/ml-ops/data/Dockerfile

@@ -0,0 +1,16 @@
+FROM python:3.12-slim-bookworm
+
+WORKDIR /app
+
+RUN apt-get update
+RUN apt-get install -y \
+    curl \
+    unzip
+
+RUN pip install --upgrade pip
+COPY requirements.txt .
+RUN pip install -r requirements.txt
+
+COPY . .
+
+CMD ["python", "main.py"]

jobs/ml-ops/data/main.py

@@ -0,0 +1,57 @@
+import boto3
+import os
+import urllib.request
+import zipfile
+
+DATA_DIR = "dataset"
+
+ZIP_URL = "http://archive.ics.uci.edu/static/public/222/bank+marketing.zip"
+ZIP_DOWNLOAD_PATH = os.path.join(DATA_DIR, "downloaded.zip")
+NESTED_ZIP_PATH = os.path.join(DATA_DIR, "bank-additional.zip")
+
+DATA_FILE = "bank-additional.csv"
+DATA_CSV_PATH = os.path.join(DATA_DIR, "bank-additional", DATA_FILE)
+
+
+def main():
+    """Pulls file from source, and uploads to a target S3 bucket"""
+
+    # Download the zip
+    os.makedirs(DATA_DIR, exist_ok=True)
+    urllib.request.urlretrieve(ZIP_URL, ZIP_DOWNLOAD_PATH)
+
+    # Extract
+    with zipfile.ZipFile(ZIP_DOWNLOAD_PATH, "r") as fh:
+        fh.extractall(DATA_DIR)
+
+    # Remove original zip
+    os.remove(ZIP_DOWNLOAD_PATH)
+
+    # Extract zips within the zip
+    with zipfile.ZipFile(NESTED_ZIP_PATH) as fh:
+        fh.extractall(DATA_DIR)
+
+    access_key = os.environ["SCW_ACCESS_KEY"]
+    secret_key = os.environ["SCW_SECRET_KEY"]
+    bucket_name = os.environ["S3_BUCKET_NAME"]
+    region_name = os.environ["SCW_REGION"]
+    s3_url = f"https://s3.{region_name}.scw.cloud"
+
+    s3 = boto3.resource(
+        "s3",
+        region_name=region_name,
+        use_ssl=True,
+        endpoint_url=s3_url,
+        aws_access_key_id=access_key,
+        aws_secret_access_key=secret_key,
+    )
+
+    bucket = s3.Bucket(name=bucket_name)
+    bucket.upload_file(
+        Filename=DATA_CSV_PATH,
+        Key=DATA_FILE,
+    )
+
+
+if __name__ == "__main__":
+    main()

jobs/ml-ops/data/requirements.txt

@@ -0,0 +1,2 @@
+boto3==1.33.2
+requests==2.31.0

jobs/ml-ops/docker-compose.yml

@@ -0,0 +1,37 @@
+version: "3"
+
+services:
+  data:
+    build:
+      context: ./data
+    depends_on:
+      - minio
+
+  training:
+    build:
+      context: ./training
+    depends_on:
+      - minio
+
+  inference:
+    build:
+      context: ./inference
+    ports:
+      - 8080:80
+    depends_on:
+      - minio
+
+  minio:
+    image: minio/minio
+    ports:
+      - "9000:9000"
+      - "9001:9001"
+    volumes:
+      - minio_storage:/data
+    environment:
+      MINIO_ROOT_USER: example
+      MINIO_ROOT_PASSWORD: example
+    command: server --console-address ":9001" /data
+
+volumes:
+  minio_storage: {}

jobs/ml-ops/containers/inference-api/Dockerfile → jobs/ml-ops/inference/Dockerfile

@@ -2,9 +2,9 @@ FROM python:3.12-slim-bookworm
 
 WORKDIR /app
 
-COPY . .
-
 RUN pip install --upgrade pip
+COPY requirements.txt .
 RUN pip install -r requirements.txt
 
-CMD ["uvicorn", "main:app", "--proxy-headers", "--host", "0.0.0.0", "--port", "80"]
+COPY . .
+CMD ["uvicorn", "main:app", "--proxy-headers", "--host", "0.0.0.0", "--port", "80"]

jobs/ml-ops/containers/inference-api/data_processing.py → jobs/ml-ops/inference/data.py

@@ -28,17 +28,21 @@ class ClientProfile(BaseModel):
     nr_employed: float
 
 
-def clean_data(data: pd.DataFrame) -> pd.DataFrame:
+def clean_profile(profile: ClientProfile) -> pd.DataFrame:
     """Removes rows with missing value(s)"""
 
-    data = data.dropna()
-    return data
+    profile_json = profile.model_dump()
+
+    cleaned = pd.DataFrame(index=[0], data=profile_json)
+    cleaned = cleaned.dropna()
+
+    return cleaned
 
 
 def transform_data(data: pd.DataFrame) -> pd.DataFrame:
     """
-    This method handles the transformation of categorical variables of the dataset into 0/1 indicators.
-    It also adds missing categorical variables that are by default false (0).
+    Transforms categorical variables of the dataset into 0/1 indicators.
+    Adds missing categorical variables that are by default false (0).
     """
 
     # # use the same category for basic education sub-categories

jobs/ml-ops/inference/example.json

@@ -0,0 +1,22 @@
+{
+  "age": 44,
+  "job": "blue-collar",
+  "marital": "married",
+  "education": "basic.4y",
+  "default": "unknown",
+  "housing": "yes",
+  "loan": "no",
+  "contact": "cellular",
+  "month": "aug",
+  "day_of_week": "thu",
+  "duration": 210,
+  "campaign": 1,
+  "pdays": 999,
+  "previous": "0",
+  "poutcome": "nonexistent",
+  "emp_var_rate": 1.4,
+  "cons_price_idx": 93.444,
+  "cons_conf_idx": -36.1,
+  "euribor3m": 4.963,
+  "nr_employed": 5228.1
+}