page_type

ms.custom

ms.contributors

languages

products

sample

team=cse

prdeb-12/21/2021

anchugh-12/21/2021

python

azure-databricks

azure-blob-storage

azure-monitor

Azure Databricks MLOps using MLflow

This is a template or sample for MLOps for Python based source code in Azure Databricks using MLflow without using MLflow Project.

This template provides the following features:

A way to run Python based MLOps without using MLflow Project, but still using MLflow for managing the end-to-end machine learning lifecycle.
Sample of machine learning source code structure along with Unit Test cases
Sample of MLOps code structure along with Unit Test cases
Demo setup to try on users subscription

Problem Summary

This demonstrates deployment scenario of Orchestrate MLOps on Azure Databricks using Databricks Notebook

Products/Technologies/Languages Used

Products & Technologies:
- Azure Databricks
- Azure Blob Storage
- Azure Monitor
Languages:
- Python

Architecture

Model Training

Batch Scoring

Individual Components

ml_experiment - sample ML experiment notebook.
ml_data - dummy data for sample model
ml_ops - sample MLOps code along with Unit Test cases, orchestrator, deployment setup.
ml_source - sample ML code along with Unit Test cases
Makefile - for build, test in local environment
requirements.txt - python dependencies

Getting Started

Prerequisites

Azure Databricks workspace
Azure Data Lake Storage Gen2 account
Visual Studio Code in local environment for development
Docker in local environment for development

Development

git clone https://github.com/Azure-Samples/azure-databricks-mlops-mlflow.git
cd azure-databricks-mlops-mlflow
Open cloned repository in Visual Studio Code Remote Container
Open a terminal in Remote Container from Visual Studio Code
make install to install sample packages (taxi_fares and taxi_fares_mlops) locally
make test to Unit Test the code locally

Package

make dist to build wheel Ml and MLOps packages (taxi_fares and taxi_fares_mlops) locally

Deployment

make databricks-deploy-code to deploy Databricks Orchestrator Notebooks, ML and MLOps Python wheel packages. If any code changes.
make databricks-deploy-jobs to deploy Databricks Jobs. If any changes in job specs.

Run training and batch scoring

To trigger training, execute make run-taxi-fares-model-training
To trigger batch scoring, execute make run-taxi-fares-batch-scoring

NOTE: for deployment and running the Databricks environment should be created first, for creating a demo environment the Demo chapter can be followed.

Observability

Check Logs, create alerts. etc. in Application Insights. Following are the few sample Kusto Query to check logs, traces, exception, etc.

Check for Error, Info, Debug Logs

Kusto Query for checking general logs for a specific MLflow experiment, filtered by mlflow_experiment_id

  traces
| extend mlflow_experiment_id = customDimensions.mlflow_experiment_id
| where timestamp > ago(30m) 
| where mlflow_experiment_id == <mlflow experiment id>
| limit 1000

Kusto Query for checking general logs for a specific Databricks job execution filtered by mlflow_experiment_id and mlflow_run_id

traces
| extend mlflow_run_id = customDimensions.mlflow_run_id
| extend mlflow_experiment_id = customDimensions.mlflow_experiment_id
| where timestamp > ago(30m) 
| where mlflow_experiment_id == <mlflow experiment id>
| where mlflow_run_id == "<mlflow run id>"
| limit 1000

Check for Exceptions

Kusto Query for checking exception log if any
```
exceptions 
| where timestamp > ago(30m)
| limit 1000
```
Check for duration of different stages in MLOps

Sample Kusto Query for checking duration of different stages in MLOps
```
dependencies 
| where timestamp > ago(30m) 
| where cloud_RoleName == 'TaxiFares_Training'
| limit 1000
```

To correlate dependencies, exceptions and traces, operation_Id can be used a filter to above Kusto Queries.

Demo

Create Databricks workspace, a storage account (Azure Data Lake Storage Gen2) and Application Insights
1. Create an Azure Account
2. Deploy resources from custom ARM template
Initialize Databricks (create cluster, base workspace, mlflow experiment, secret scope)
1. Get Databricks CLI Host and Token
2. Authenticate Databricks CLI make databricks-authenticate
3. Execute make databricks-init
Create Azure Data Lake Storage Gen2 Container and upload data
1. Create Azure Data Lake Storage Gen2 Container named - taxifares
2. Upload as blob taxi-fares data files into Azure Data Lake Storage Gen2 container named - taxifares
Put secrets to Mount ADLS Gen2 Storage using Shared Access Key
1. Get Azure Data Lake Storage Gen2 account name created in step 1
2. Get Shared Key for Azure Data Lake Storage Gen2 account
3. Execute make databricks-secrets-put to put secret in Databricks secret scope
Put Application Insights Key as a secret in Databricks secret scope (optional)
1. Get Application Insights Key created in step 1
2. Execute make databricks-add-app-insights-key to put secret in Databricks secret scope
Package and deploy into Databricks (Databricks Jobs, Orchestrator Notebooks, ML and MLOps Python wheel packages)
1. Execute make deploy
Run Databricks Jobs
1. To trigger training, execute make run-taxifares-model-training
2. To trigger batch scoring, execute make run-taxifares-batch-scoring
Expected results
1. Azure resources
2. Databricks jobs
3. Databricks mlflow experiment
4. Databricks mlflow model registry
5. Output of batch scoring

Additional Details

Related resources

Glossaries

Application developer : It is a role that work mainly towards operationalize of machine learning.
Data scientist : It is a role to perform the data science parts of the project

Azure-Samples/azure-databricks-mlops-mlflow