Quickstart

Video-tutorial

You can go through the written quickstart here or watch the video on YouTube:

---

Prerequisites

Before you start, make sure that you have the following resources created in Azure and have their names ready to input to the plugin:

• Azure Subscription ID

• Azure Resource Group

• Azure ML workspace

• Azure ML Compute Cluster

Depending on the type of flow you want to use, you might also need: - Azure Storage Account and Storage Container - Azure Storage Key (will be used to execute the pipeline) - Azure Container Registry

Project initialization

1. Make sure that you’re logged into Azure (az login).

2. Prepare new virtual environment with Python >=3.8. Install the packages

   pip install "kedro>=0.18.5,<0.19" "kedro-docker" "kedro-azureml"
   

2. Create new project (e.g. from starter)

   kedro new --starter=spaceflights
   
   Project Name
   ============
   Please enter a human readable name for your new project.
   Spaces, hyphens, and underscores are allowed.
    [Spaceflights]: kedro_azureml_demo
   
   The project name 'kedro_azureml_demo' has been applied to:
   - The project title in /Users/marcin/Dev/tmp/kedro-azureml-demo/README.md
   - The folder created for your project in /Users/marcin/Dev/tmp/kedro-azureml-demo
   - The project's python package in /Users/marcin/Dev/tmp/kedro-azureml-demo/src/kedro_azureml_demo
   

3. Go to the project’s directory: cd kedro-azureml-demo

4. Add kedro-azureml to src/requirements.txt

5. (optional) Remove kedro-telemetry from src/requirements.txt or set appropriate settings (https://github.com/kedro-org/kedro-plugins/tree/main/kedro-telemetry).

6. Install the requirements pip install -r src/requirements.txt

7. Initialize Kedro Azure ML plugin, it requires the Azure resource names as stated above. Experiment name can be anything you like (as long as it’s allowed by Azure ML).

   There are two options, which determine how you should initialize the plugin (don’t worry, you can change it later 👍 ):

   1. Use docker image flow (shown in the Quickstart video) - more suitable for MLOps processes with better experiment repeatability guarantees

   2. Use code upload flow - more suitable for Data Scientists’ fast experimentation and pipeline development

Usage: kedro azureml init [OPTIONS] SUBSCRIPTION_ID RESOURCE_GROUP
                          WORKSPACE_NAME EXPERIMENT_NAME CLUSTER_NAME

  Creates basic configuration for Kedro AzureML plugin

Options:
  --azureml-environment, --aml-env TEXT
                                  Azure ML environment to use with code flow
  -d, --docker-image TEXT         Docker image to use
  -a, --storage-account-name TEXT
                                  Name of the storage account (if you want to
                                  use Azure Blob Storage for temporary data)
  -c, --storage-container TEXT    Name of the storage container (if you want
                                  to use Azure Blob Storage for temporary
                                  data)
  --use-pipeline-data-passing     (flag) Set, to use EXPERIMENTAL pipeline
                                  data passing

For docker image flow (1.), use the following init command:

kedro azureml init <AZURE_SUBSCRIPTION_ID> <AZURE_RESOURCE_GROUP> <AML_WORKSPACE_NAME> <EXPERIMENT_NAME> <COMPUTE_NAME> \
 --docker-image <YOUR_ARC>.azurecr.io/<IMAGE_NAME>:latest -a <STORAGE_ACCOUNT_NAME> -c <STORAGE_CONTAINER_NAME>

For code upload flow (2.), use the following init command:

kedro azureml init <AZURE_SUBSCRIPTION_ID> <AZURE_RESOURCE_GROUP> <AML_WORKSPACE_NAME> <EXPERIMENT_NAME> <COMPUTE_NAME> \
 --aml-env <YOUR_ARC>.azurecr.io/<IMAGE_NAME>:latest -a <STORAGE_ACCOUNT_NAME> -c <STORAGE_CONTAINER_NAME>

Note

If you want to pass data between nodes using the built-in Azure ML pipeline data passing, specify option --use-pipeline-data-passing instead of -a and -c options.

Note that pipeline data passing feature is experimental 🧑‍🔬 See Azure Data Assets for more information about this.

Adjusting the Data Catalog

8. Adjust the Data Catalog - the default one stores all data locally, whereas the plugin will automatically use Azure Blob Storage / Azure ML built-in storage (if pipeline data passing was enabled). Only input data is required to be read locally.

   Final conf/base/catalog.yml should look like this:

   companies:
     type: pandas.CSVDataset
     filepath: data/01_raw/companies.csv
     layer: raw
   
   reviews:
     type: pandas.CSVDataset
     filepath: data/01_raw/reviews.csv
     layer: raw
   
   shuttles:
     type: pandas.ExcelDataset
     filepath: data/01_raw/shuttles.xlsx
     layer: raw
   

Pick your deployment option

For the project’s code to run on Azure ML it needs to have an environment with the necessary dependencies.

9. Start by executing the following command:

   kedro docker init
   

   This command creates a several files, including Dockerfile and .dockerignore. These can be adjusted to match the workflow for your project.

Depending on whether you want to use code upload when submitting an experiment or not, you would need to add the code and any possible input data to the Docker image.

(Option 1) Docker image flow

This option is also shown in the video-tutorial above.

Note

Note that using docker image flow means that every time you change your pipeline’s code,

you will need to build and push the docker image to ACR again.

We recommend this option for CI/CD-automated MLOps workflows.

10. Ensure that in the azureml.yml you have azure.code_directory set to null, and docker.image is filled:

    azure:
      code_directory: ~
    # rest of the azureml.yml file
    docker:
       image: your-container-registry.azurecr.io/kedro-azureml:latest
    

11. Adjust the .dockerignore file to include any other files to be added to the Docker image, such as !data/01_raw for the raw data files.

12. Invoke docker build:

    kedro docker build --docker-args "--build-arg=BASE_IMAGE=python:3.9" --image=<image tag from conf/base/azureml.yml>
    

13. Once finished, login to ACR:

    az acr login --name <acr repo name>
    

    and push the image:

    docker push <image tag from conf/base/azureml.yml>
    

(Option 2) Code upload flow

10. Everything apart from the section install project requirements can be removed from the Dockerfile.

This plugin automatically creates empty .amlignore file (see the official docs) which means that all of the files (including potentially sensitive ones!) will be uploaded to Azure ML. Modify this file if needed.

See example Dockerfile for code upload flow

ARG BASE_IMAGE=python:3.9
FROM $BASE_IMAGE

# install project requirements
COPY src/requirements.txt /tmp/requirements.txt
RUN pip install -r /tmp/requirements.txt && rm -f /tmp/requirements.txt

11. Ensure azure.code_directory: "." is set in the azureml.yml config file (it’s set if you’ve used --aml_env during init above).

12. Build the image:

    kedro docker build --docker-args "--build-arg=BASE_IMAGE=python:3.9" --image=<acr repo name>.azurecr.io/kedro-base-image:latest
    

12. Login to ACR and push the image:

    az acr login --name <acr repo name>
    docker push <acr repo name>.azurecr.io/kedro-base-image:latest
    

13. Register the Azure ML Environment:

    az ml environment create --name <environment-name> --image <acr repo name>.azurecr.io/kedro-base-image:latest
    

Now you can re-use this environment and run the pipeline without the need to build the docker image again (unless you add some dependencies to your environment, obviously 😉 ).

Warning

Azure Code upload feature has issues with empty folders as identified in GitHub #33, where empty folders or folders with empty files might not get uploaded to Azure ML, which might result in the failing pipeline.

We recommend to:

- make sure that Kedro environments you intent to use in Azure have at least one non-empty file specified

- gracefully handle folder creation in your pipeline’s code (e.g. if your code depends on an existence of some folder)

The plugin will do it’s best to handle some of the edge-cases, but the fact that some of your files might not be captured by Azure ML SDK is out of our reach.

Run the pipeline

14. Run the pipeline on Azure ML Pipelines. Here, the Azure Subscription ID and Storage Account Key will be used:

    kedro azureml run
    

    If you’re using Azure Blob Storage for temporary data (-a, -c options during init), you will most likely see the following prompt:

    Environment variable AZURE_STORAGE_ACCOUNT_KEY not set, falling back to CLI prompt
    Please provide Azure Storage Account Key for storage account <azure-storage-account>:
    

    Input the storage account key and press [ENTER] (input will be hidden).

    If you’re using pipeline data passing (--use-pipeline-data-passing option during init), you’re already set.

11. Plugin will verify the configuration (e.g. the existence of the compute cluster) and then it will create a Job in the Azure ML. The URL to view the job will be displayed in the console output.

12. (optional) You can also use
    kedro azureml run -s <azure-subscription-id> --wait-for-completion
    to actively wait for the job to finish. Execution logs will be streamed to the console.

    RunId: placid_pot_bdcyntnkvn
    Web View: https://ml.azure.com/runs/placid_pot_bdcyntnkvn?wsid=/subscriptions/<redacted>/resourcegroups/<redacted>/workspaces/ml-ops-sandbox
    
    Streaming logs/azureml/executionlogs.txt
    ========================================
    
    [2022-07-22 11:45:38Z] Submitting 2 runs, first five are: 1ee5f43f:8cf2e387-e7ec-44cc-9615-2108891153f7,7d81aeeb:c8b837a9-1f79-4971-aae3-3191b29b42e8
    [2022-07-22 11:47:02Z] Completing processing run id c8b837a9-1f79-4971-aae3-3191b29b42e8.
    [2022-07-22 11:47:25Z] Completing processing run id 8cf2e387-e7ec-44cc-9615-2108891153f7.
    [2022-07-22 11:47:26Z] Submitting 1 runs, first five are: 362b9632:7867ead0-b308-49df-95ca-efa26f8583cb
    [2022-07-22 11:49:27Z] Completing processing run id 7867ead0-b308-49df-95ca-efa26f8583cb.
    [2022-07-22 11:49:28Z] Submitting 2 runs, first five are: 03b2293e:e9e210e7-10ab-4010-91f6-4a40aabf3a30,4f9ccafb:3c00e735-cd3f-40c7-9c1d-fe53349ca8bc
    [2022-07-22 11:50:50Z] Completing processing run id e9e210e7-10ab-4010-91f6-4a40aabf3a30.
    [2022-07-22 11:50:51Z] Submitting 1 runs, first five are: 7a88df7a:c95c1488-5f55-48fa-80ce-971d5412f0fb
    [2022-07-22 11:51:26Z] Completing processing run id 3c00e735-cd3f-40c7-9c1d-fe53349ca8bc.
    [2022-07-22 11:51:26Z] Submitting 1 runs, first five are: a79effc8:0828c39a-6f02-43f5-acfd-33543f0d6c74
    [2022-07-22 11:52:38Z] Completing processing run id c95c1488-5f55-48fa-80ce-971d5412f0fb.
    [2022-07-22 11:52:39Z] Submitting 1 runs, first five are: 0a18d6d6:cb9c8f61-e129-4394-a795-ab70be74eb0f
    [2022-07-22 11:53:03Z] Completing processing run id 0828c39a-6f02-43f5-acfd-33543f0d6c74.
    [2022-07-22 11:53:04Z] Submitting 1 runs, first five are: 1af5c8de:2821dc44-3399-4a26-9cdf-1e8f5b7d6b62
    [2022-07-22 11:53:28Z] Completing processing run id cb9c8f61-e129-4394-a795-ab70be74eb0f.
    [2022-07-22 11:53:51Z] Completing processing run id 2821dc44-3399-4a26-9cdf-1e8f5b7d6b62.
    
    Execution Summary
    =================
    RunId: placid_pot_bdcyntnkvn
    

|Kedro AzureML Pipeline execution|

---

Using a different compute cluster for specific nodes

For certain nodes it can make sense to run them on a different compute clusters (e.g. High Memory or GPU). This can be achieved using Node tags and adding additional compute targets in your azureml.yml.

After creating an additional compute cluster in your AzureML workspace, in this case the additional cluster is called cpu-cluster-8, we can add it in our azureml.yml under an alias (in this case chunky).

compute:
  __default__:
    cluster_name: "cpu-cluster"
  chunky:
    cluster_name: "cpu-cluster-8"

Now we are able to reference this compute target in our kedro pipelines using kedro node tags:

[
    node(
        func=preprocess_companies,
        inputs="companies",
        outputs="preprocessed_companies",
        name="preprocess_companies_node",
        tags=["chunky"]
    ),
    node(
        func=preprocess_shuttles,
        inputs="shuttles",
        outputs="preprocessed_shuttles",
        name="preprocess_shuttles_node",
    ),
    node(
        func=create_model_input_table,
        inputs=["preprocessed_shuttles", "preprocessed_companies", "reviews"],
        outputs="model_input_table",
        name="create_model_input_table_node",
        tags=["chunky"]
    ),
],

When running our project, preprocess_companies and create_model_input_table will be run on cpu-cluster-8 while all other nodes are run on the default cpu-cluster.

Marking a node as deterministic

By default the plugin will mark all nodes of the Azure ML pipeline as non-deterministic, which means that Azure ML will not reuse the results of the node if it has been run before.

To mark a node as deterministic, you can simply add the deterministic tag to the node. This also implies the tag is reserved and cannot be used for compute types.

Distributed training

The plugins supports distributed training via native Azure ML distributed orchestration, which includes:

• MPI - https://learn.microsoft.com/en-us/azure/machine-learning/how-to-train-distributed-gpu#mpi

• PyTorch - https://learn.microsoft.com/en-us/azure/machine-learning/how-to-train-distributed-gpu#pytorch

• TensorFlow - https://learn.microsoft.com/en-us/azure/machine-learning/how-to-train-distributed-gpu#tensorflow

If one of your Kedro’s pipeline nodes requires distributed training (e.g. you train a neural network with PyTorch), you can mark the node with distributed_job decorator from kedro_azureml.distributed.decorators and use native Kedro parameters to specify the number of nodes you want to spawn for the job. An example for PyTorch looks like this:

#                    | use appropriate framework
#                   \|/                      \/ specify the number of distributed nodes to spawn for the job
@distributed_job(Framework.PyTorch, num_nodes="params:num_nodes")
def train_model_pytorch(
    X_train: pd.DataFrame, y_train: pd.Series, num_nodes: int, max_epochs: int
):
    # rest of the code
    pass

In the pipeline you would use this node like that:

node(
    func=train_model_pytorch,
    inputs=["X_train", "y_train", "params:num_nodes", "params:max_epochs"],
    outputs="regressor",
    name="train_model_node",
),

and that’s it! The params: you use support namespacing as well as overriding at runtime, e.g. when launching the Azure ML job:

kedro azureml run -s <subscription id> --params '{"data_science": {"active_modelling_pipeline": {"num_nodes": 4}}}'

The distributed_job decorator also supports “hard-coded” values for number of nodes:

@distributed_job(Framework.PyTorch, num_nodes=2) # no need to use Kedro params here
def train_model_pytorch(
    X_train: pd.DataFrame, y_train: pd.Series, num_nodes: int, max_epochs: int
):
    # rest of the code
    pass

We have tested the implementation heavily with PyTorch (+PyTorch Lightning) and GPUs. If you encounter any problems, drop us an issue on GitHub!

Run customization

In case you need to customize pipeline run context, modifying configuration files is not always the most convinient option. Therefore, kedro azureml run command provides a few additional options you may find useful:

• --subscription_id overrides Azure Subscription ID,

• --azureml_environment overrides the configured Azure ML Environment,

• --image modifies the Docker image used during the execution,

• --pipeline allows to select a pipeline to run (by default, the __default__ pipeline is started),

• --params takes a JSON string with parameters override (JSONed version of conf/*/parameters.yml, not the Kedro’s params: syntax),

• --env-var KEY=VALUE sets the OS environment variable injected to the steps during runtime (can be used multiple times).

• --load-versions specifies a particular dataset version (timestamp) for loading (similar behavior as Kedro)

• --on-job-scheduled  path.to.module:my_function specifies a callback function to be called on the azureml pipeline job start (example below)

# src/mymodule/myfile.py
def save_output_callback(job):
    """saves the pipeline job name to a file"""
    with open("myfile.txt", "w") as f:
        f.write(job.name)

kedro azureml run --on-job-scheduled mymodule.myfile:save_output_callback