ccw2145/mlops_stacks_demo

MLOps Stacks demo

mlops_stacks_demo

This directory contains an ML project based on the default Databricks MLOps Stacks, defining a production-grade ML pipeline for automated retraining and batch inference of an ML model on tabular data.

See the full pipeline structure below. The MLOps Stacks README contains additional details on how ML pipelines are tested and deployed across each of the dev, staging, prod environments below.

Code structure

This project contains the following components:

Component	Description
ML Code	Example ML project code, with unit tested Python modules and notebooks
ML Assets as Code	ML pipeline assets (training and batch inference jobs with schedules, etc) configured and deployed through databricks CLI bundles
CI/CD	GitHub Actions workflows to test and deploy ML code and assets

contained in the following files:

mlops_stacks_demo        <- Root directory. Both monorepo and polyrepo are supported.
│
├── mlops_stacks_jll       <- Contains python code, notebooks and ML assets related to one ML project. 
│   │
│   ├── requirements.txt        <- Specifies Python dependencies for ML code (for example: model training, batch inference).
│   │
│   ├── databricks.yml          <- databricks.yml is the root bundle file for the ML project that can be loaded by databricks CLI bundles. It defines the bundle name, workspace URL and asset config component to be included.
│   │
│   ├── training                <- Training folder contains Notebook that trains and registers the model with feature store support.
│   │
│   ├── feature_engineering     <- Feature computation code (Python modules) that implements the feature transforms.
│   │                              The output of these transforms get persisted as Feature Store tables. Most development
│   │                              work happens here.
│   │
│   ├── validation              <- Optional model validation step before deploying a model.
│   │
│   ├── monitoring              <- Model monitoring, feature monitoring, etc.
│   │
│   ├── deployment              <- Deployment and Batch inference workflows
│   │   │
│   │   ├── batch_inference     <- Batch inference code that will run as part of scheduled workflow.
│   │   │
│   │   ├── model_deployment    <- As part of CD workflow, deploy the registered model by assigning it the appropriate alias.
│   │
│   │
│   ├── tests                   <- Unit tests for the ML project, including the modules under `features`.
│   │
│   ├── assets               <- ML asset (ML jobs, MLflow models) config definitions expressed as code, across dev/staging/prod/test.
│       │
│       ├── model-workflow-asset.yml                <- ML asset config definition for model training, validation, deployment workflow
│       │
│       ├── batch-inference-workflow-asset.yml      <- ML asset config definition for batch inference workflow
│       │
│       ├── feature-engineering-workflow-asset.yml  <- ML asset config definition for feature engineering workflow
│       │
│       ├── ml-artifacts-asset.yml                  <- ML asset config definition for model and experiment
│       │
│       ├── monitoring-workflow-asset.yml           <- ML asset config definition for data monitoring workflow
│
├── .github                     <- Configuration folder for CI/CD using GitHub Actions. The CI/CD workflows deploy ML assets defined in the `./assets/*` folder with databricks CLI bundles.

Using this repo

The table below links to detailed docs explaining how to use this repo for different use cases.

This project comes with example ML code to train, validate and deploy a regression model to predict NYC taxi fares. If you're a data scientist just getting started with this repo for a brand new ML project, we recommend adapting the provided example code to your ML problem. Then making and testing ML code changes on Databricks or your local machine. Follow the instructions from the ML quickstart.

When you're satisfied with initial ML experimentation (e.g. validated that a model with reasonable performance can be trained on your dataset) and ready to deploy production training/inference pipelines, ask your ops team to follow the MLOps setup guide to configure CI/CD and deploy production ML pipelines.

After that, follow the ML pull request guide and ML asset config guide to propose, test, and deploy changes to production ML code (e.g. update model parameters) or pipeline assets (e.g. use a larger instance type for model training) via pull request.

Role	Goal	Docs
Data Scientist	Get started writing ML code for a brand new project	ML quickstart.
MLOps / DevOps	Set up CI/CD for the current ML project	MLOps setup guide
Data Scientist	Update production ML code (e.g. model training logic) for an existing project	ML pull request guide
Data Scientist	Modify production model ML assets, e.g. model training or inference jobs	ML asset config guide

Monorepo

It's possible to use the repo as a monorepo that contains multiple projects. All projects share the same workspaces and service principals.

For example, assuming there's existing repo with root directory name monorepo_root_dir and project name project1

1. Create another project from databricks bundle init with project name project2 and root directory name project2.
2. Copy the internal directory project2/project2 to root directory of existing repo monorepo_root_dir/project2.
3. Copy yaml files from project2/.github/workflows/ to monorepo_root_dir/.github/workflows/ and make sure there's no name conflicts.