Anomaly Detection Pipeline on Azure Databricks

The following is an anomaly detection data pipeline on Azure Databricks. This solution was built to demonstrate how to build Advance Analytics Pipelines on Azure Databricks, with a particular focus on the Spark MLLib library. This solution includes:

1. Initial ETL Data loading process into SparkSQL tables
2. Model training and scoring
   • Explanation of Pipelines, Transformer and Estimators
   • Sample Custom Estimator (PCAAnomaly)
3. Persisting trained models
4. Productionizing models through
   • Batch inference
   • Streaming

Architecture

Data

KDD Cup 1999 Data

Deployment

• Ensure you are in the root of the repository
• To deploy the solution, use one of the following commands:
  1. (Easiest) Using pre-built docker container: docker run -it devlace/azdatabricksanomaly
  2. Build and run the container locally: make deploy_w_docker
  3. Deploy using local environment (see requirements below): make deploy
• Follow the prompts to login to Azure, name of resource group, deployment location, etc.
• When prompted for a Databricks Host, enter the full name of your databricks workspace host, e.g. https://southeastasia.azuredatabricks.net
• When prompted for a token, you can generate a new token in the databricks workspace.

To view additional make commands run make

For local deployment

Requirements

• Azure CLI 2.0+
• Python virtualenv or Anaconda
• jq tool
• Check the requirements.txt for list of necessary Python packages. (will be installed by make requirements)

Development environment

• The following works with Windows Subsystem for Linux
• Clone this repository
• cd azure-databricks-anomaly
• Create a python environment (Virtualenv or Conda). The following uses virtualenv.
  • virtualenv . This creates a python virtual environment to work in.
  • source bin/activate This activates the virtual environment.
• make requirements. This installs python dependencies in the virtual environment.

Project Organization

---

├── LICENSE
├── Makefile           <- Makefile with commands like `make data` or `make train`
├── README.md          <- The top-level README for developers using this project.
├── deploy             <- Deployment artifacts
│   │
│   └── databricks     <- Deployment artifacts in relation to the Databricks workspace
│   │
│   └── deploy.sh      <- Deployment script to deploy all Azure Resources
│   │
│   └── azuredeploy.json <- Azure ARM template w/ .parameters file
│   │
│   └── Dockerfile     <- Dockerfile for deployment
│
├── models             <- Trained and serialized models, model predictions, or model summaries
│
├── notebooks          <- Jupyter notebooks. Naming convention is a number (for ordering),
│                         the creator's initials, and a short `-` delimited description, e.g.
│                         `1.0-jqp-initial-data-exploration`.
│
├── references         <- Contains the powerpoint presentation, and other reference materials.
│
├── requirements.txt   <- The requirements file for reproducing the analysis environment, e.g.
│                         generated with `pip freeze > requirements.txt`
│
├── setup.py           <- makes project pip installable (pip install -e .) so src can be imported

---

Project based on the cookiecutter data science project template. #cookiecutterdatascience