Skip to content

Source: examples/ml-lifecycle/README.md — this page is rendered live from that file.

ML Lifecycle Example — Loan Default Classifier

Examples > ML Lifecycle

TL;DR — End-to-end, reproducible Azure ML pipeline: deterministic synthetic data → dbt feature engineering → scikit-learn training with AUC >= 0.70 → managed online endpoint → data-drift monitoring. Works offline for tests; documented for Azure-side deployment.

Table of Contents

Architecture

graph LR
    subgraph Data["Data"]
        GEN["Deterministic<br/>Loan Generator<br/>(numpy seed=42)"]
        CSV["loans.csv<br/>(sha256-stable)"]
        DBT["dbt Bronze/Silver/Gold<br/>(feature engineering)"]
    end

    subgraph Training["Training"]
        SPLIT["Stratified<br/>Train/Test 75/25"]
        FEAT["ColumnTransformer<br/>(scale + one-hot)"]
        LR["LogisticRegression<br/>(sklearn)"]
        AUC{"AUC >= 0.70?"}
        MODEL["model.pkl<br/>+ metrics.json"]
    end

    subgraph Deploy["Deploy (Azure)"]
        AML["Azure ML<br/>Workspace"]
        REG["Model<br/>Registry"]
        EP["Managed<br/>Online Endpoint"]
    end

    subgraph Monitor["Monitor"]
        PROD["Production<br/>Scoring Inputs"]
        DRIFT["Evidently / PSI<br/>Drift Detector"]
        ALERT["Alert<br/>(> threshold)"]
    end

    GEN --> CSV --> DBT --> SPLIT --> FEAT --> LR --> AUC
    AUC -->|yes| MODEL
    AUC -->|no| LR
    MODEL --> REG --> EP
    EP --> PROD --> DRIFT
    DRIFT -->|drift > threshold| ALERT
    ALERT -.retrain.-> GEN

See ARCHITECTURE.md for detailed component breakdown.

Prerequisites

Tool Version Notes
Python 3.10+ 3.12 tested
pip latest
Azure CLI 2.56+ only for Azure deployment
az ml extension 2.x az extension add -n ml
dbt-core + dbt-duckdb 1.7+ only for dbt parse/compile

Python packages (installed locally or via the ml extra group):

  • numpy, pandas, scikit-learn>=1.4, joblib
  • evidently (optional — PSI fallback works without it)
  • azure-ai-ml, azure-identity (optional — only for az ml automation)

Cost Estimate

Component Est. Monthly (USD)
Azure ML workspace (Basic) $0
Storage (ADLS Gen2, ~10 GB) ~$0.50
Key Vault (standard, ~1k ops) ~$0.05
Application Insights (low-volume) ~$5
Managed online endpoint (1x Standard_DS3_v2, 24/7) ~$140
Total (dev/demo, DS3 always-on) ~$145

Cut ~75% by stopping the endpoint when idle (az ml online-endpoint update ... --traffic blue=0) or deploying to a serverless compute option when available.

Directory Structure

examples/ml-lifecycle/
  README.md                    (this file)
  ARCHITECTURE.md              (component diagram + flow)
  data/
    generators/
      generate_loan_data.py    (seeded synthetic CSV)
      tests/
        test_generate_loan_data.py
  domains/
    dbt/
      dbt_project.yml
      models/
        schema.yml             (sources + docs)
        bronze/brz_loan_applications.sql
        silver/slv_loan_features.sql
        gold/gld_training_features.sql
      seeds/
        loan_applications.csv  (small deterministic seed for CI)
  training/
    train.py                   (sklearn pipeline, AUC >= 0.70 guarantee)
    score.py                   (Azure ML online endpoint handler)
    conda.yaml
    mlproject.yaml             (Azure ML pipeline spec)
    tests/
      test_train.py            (AUC assertion end-to-end)
      test_score.py
  deploy/
    bicep/
      main.bicep               (AML + Storage + KV + AppI + endpoint)
      params.dev.json
    register_model.sh
    deploy_endpoint.sh
  monitoring/
    drift_detection.py         (Evidently + PSI fallback)
    tests/
      test_drift_detection.py
  contracts/
    loan_training_features.yaml
    loan_prediction_contract.yaml

Quick Start (Offline)

Everything below runs without Azure, using only the synthetic generator.

Generate deterministic data

python examples/ml-lifecycle/data/generators/generate_loan_data.py \
  --rows 5000 --seed 42 \
  --output outputs/loans.csv

Same seed → same SHA-256.

Train

python examples/ml-lifecycle/training/train.py \
  --rows 5000 --seed 42 \
  --output-dir outputs/ml
# → AUC=0.84, accuracy=0.82, model.pkl + metrics.json written.

Run tests

pytest examples/ml-lifecycle/ -v

Full Azure Deployment

# 1. Provision AML workspace + endpoint shell
az deployment group create \
  --resource-group csa-mllife-rg \
  --template-file examples/ml-lifecycle/deploy/bicep/main.bicep \
  --parameters @examples/ml-lifecycle/deploy/bicep/params.dev.json

# 2. Train locally
python examples/ml-lifecycle/training/train.py --output-dir examples/ml-lifecycle/training/outputs

# 3. Register the model
./examples/ml-lifecycle/deploy/register_model.sh \
  csa-mllife-rg csamllife-aml-dev

# 4. Deploy to the endpoint
./examples/ml-lifecycle/deploy/deploy_endpoint.sh \
  csa-mllife-rg csamllife-aml-dev csamllife-loan-default-dev

# 5. Smoke test
az ml online-endpoint invoke \
  --name csamllife-loan-default-dev \
  --request-file examples/ml-lifecycle/training/tests/sample_request.json

Training

See training/train.py.

  • Stratified 75/25 split (seeded).
  • ColumnTransformer pipeline: StandardScaler for numerics, OneHotEncoder(handle_unknown="ignore") for categoricals.
  • LogisticRegression(max_iter=500, solver="lbfgs").
  • Writes model.pkl (joblib) and metrics.json.

Contract: test_train.py asserts AUC >= 0.70. Typical AUC on the synthetic data is ~0.84.

Scoring

See training/score.py.

Request (POST body): 

{
  "data": [
    {"applicant_age": 35, "annual_income": 72500, ...}
  ]
}

Response: 

{
  "predictions": [
    {"probability_default": 0.18, "prediction": 0}
  ]
}

Same handler works standalone (import score_records) and inside the Azure ML inference server (via init() + run()).

Drift Monitoring

See monitoring/drift_detection.py.

  • Preferred backend: Evidently (DataDriftPreset).
  • Fallback: Population Stability Index (PSI) — pure NumPy, zero extra deps.
python examples/ml-lifecycle/monitoring/drift_detection.py \
  --reference outputs/train_snapshot.csv \
  --current   outputs/prod_snapshot.csv
# Exit code 0 = no drift, 1 = drift detected.

Schedule via cron / Container Apps Job / Data Factory. The contract in contracts/loan_prediction_contract.yaml specifies a 6-hour cadence by default.

Teardown

az group delete --name csa-mllife-rg --yes --no-wait

Testing

  • Offline (no Azure): 
    pytest examples/ml-lifecycle/ -v
  • dbt parse + compile: 
    cd examples/ml-lifecycle/domains/dbt
dbt parse --profiles-dir <stub-profiles> --target ci
dbt compile --profiles-dir <stub-profiles> --target ci
    The stub profile matches .github/workflows/dbt-ci.yml — csa_ml_lifecycle is already in that workflow's profile block.

Contracts