A4 - architecture

• normally, a ML training architecture includes:

project/
├── train.py
├── data.csv
├── requirements.txt
└── model.pkl

# Run
>>> python train.py

• for large scare projects, Azure provides a much better architecture:

Azure subscription
      │
      ▼
Resource Group
      │
      ▼
Azure ML Workspace
      │
      ├── Data assets
      ├── Environments
      ├── Jobs
      ├── Experiments
      ├── Models
      ├── Components
      ├── Pipelines
      ├── Compute
      └── Endpoints

• the workspace itself is a resource of type Microsoft.MachineLearningServices/workspaces

• it is not a VM doing the training, but it manages and references compute, which executes the training job

• the workspace does not store the assets either, rather it integrates with storage and tracks metadata/references/artifacts around these assets

• eg: workspace has the metadata/reference for the data asset detector-data:v3, which points to Azure storage containing actual data bytes

• workspace traditionally contains:

  • storage for datasets/artifacts, outputs, logs, model files
  • key vault for secure storage of secret/keys
  • container registry to store container images used for ML environments/deployments
  • monitoring services for operational telemetry and observability

• they are peer resources in the resource group

Subscription
    │
    ▼
rg-ai300-dev
    ├── mlw-ai300-dev        ← Azure ML workspace
    ├── stai300dev               ← Storage
    ├── kv-ai300-dev           ← Key Vault
	└── cr-ai300dev             ← Container Registry

• ML Studio is a web interface to interact with the workspace, useful for:

  • inspecting jobs
  • exploring assets
  • configuring resources
  • monitoring experiments
  • viewing models
  • managing endpoints
  • interactive work

• great for learning/exploration

• SDK can be used to interact programmatically from python, useful to:

  • submit jobs
  • create environments
  • manage data assets
  • register models
  • create endpoints
  • build pipelines

• great for programmatic ML workflows

from azure.ai.ml import MLClient

ml_client = MLClient(
	# identity/authentication
    credential=...,
	# target resource
    subscription_id=...,
    resource_group_name=...,
    workspace_name=...
)


from azure.identity import DefaultAzureCredential

• DefaultAzureCredential tries supported authentication mechanisms in a defined chain so that the same application can work in different environments

• application code does not need password

• CLI can also be used to interact from the terminal/CI

• great for automation/CI/CD

• environments define the runtime dependencies used by jobs/deployments

• training code + environment + compute -> training job

• a job represents work submitted to Azure ML

• eg: run train.py using environment X on compute Y with data Z

• a job is a particular execution of ML work whereas an experiment groups/organises related runs conceptually

• eg:

Experiment: detector-failure-model
│
├── Job/run 001
│   learning rate = .01
│
├── Job/run 002
│   learning rate = .001
│
└── Job/run 003
    learning rate = .0001

• after training, the models can be registered and tracked along with metadata/lineage

# resource-management view
RESOURCE GROUP
├── Azure ML Workspace
├── Storage
├── Key Vault
└── Container Registry

# ML lifecycle view
AZURE ML WORKSPACE
├── jobs
├── models
├── data assets
├── environments
├── components
└── endpoints