Implementation:Interpretml Interpret Powerlift RunAzureVM

Knowledge Sources	Interpretml Interpret
Domains	Benchmarking, Cloud_Infrastructure, Azure, VM_Orchestration
Last Updated	2026-02-07 12:00 GMT

Overview

Azure VM provisioning runner that creates, configures, and manages the full lifecycle of Azure Virtual Machines for executing Powerlift benchmark trials, including networking, SSH key generation, startup scripting, and RBAC permission assignment.

Description

This module implements the remote process logic invoked by the AzureVMInstance executor. It contains two main functions:

assign_contributor_permissions() -- Manages Azure RBAC role assignments for VMs using a priority queue pattern:

• Uses a min-heap to process VMs in order of creation time.
• Waits for VM creation to complete before querying the VM's managed identity principal ID.
• Assigns the Contributor role to both the VM's managed identity and the client user.
• Retries on HttpResponseError, recreating API clients as needed.

run_azure_process() -- Orchestrates the full VM lifecycle:

• Generates an SSH key pair using cryptography.hazmat for secure VM access.
• Creates a virtual network and subnet (named powerlift-vnet-{location} and powerlift-subnet-{location}).
• For each runner instance, creates a network interface and provisions a VM with:
  • Configurable size, OS image (publisher/offer/SKU/version), and disk type.
  • System-assigned managed identity for Azure resource access.
  • A base64-encoded custom data startup script injected via cloud-init.
• The startup script installs Python, creates a virtual environment, retrieves experiment configuration from the database via psql, installs dependencies, and runs the experiment in a loop.
• VMs self-delete on completion using Azure CLI (az vm delete).
• Network interfaces and OS disks have deleteOption: Delete for automatic cleanup.

Unlike the ACI runner, this module does not poll for container termination since VMs manage their own lifecycle through the self-delete mechanism in the startup script.

Usage

This module is not called directly by users. It is invoked by the AzureVMInstance executor via multiprocessing.Pool.apply_async(). Understanding this module is essential for debugging Azure VM provisioning or customizing VM startup behavior.

Code Reference

Source Location

• Repository: Interpretml_Interpret
• File: python/powerlift/powerlift/run/azure_vm.py

Signature

def assign_contributor_permissions(
    compute_client,
    auth_client,
    max_undead,
    credential,
    subscription_id,
    client_id,
    resource_group_name,
    resource_uris,
    vms,
):
    ...

def run_azure_process(
    experiment_id,
    n_instances,
    uri,
    resource_uris,
    timeout,
    azure_json,
    credential,
    location,
    vm_size,
    image_publisher,
    image_offer,
    image_sku,
    image_version,
    disk_type,
    max_undead,
    delete_on_complete,
    batch_id,
):
    ...

Import

from powerlift.run.azure_vm import run_azure_process, assign_contributor_permissions

I/O Contract

Inputs

Name	Type	Required	Description
experiment_id	int	Yes	Experiment identifier to run trials for
n_instances	int	Yes	Number of VMs to create
uri	str	Yes	Database connection URI (embedded in startup script)
resource_uris	List[str]	No	Azure resource URIs to grant contributor permissions
timeout	float	Yes	Timeout in seconds per trial execution
azure_json	dict	Yes	Azure credentials (tenant_id, client_id, client_secret, subscription_id, resource_group)
credential	object	No	Pre-built Azure credential object
location	str	No	Azure region; defaults to resource group location
vm_size	str	Yes	Azure VM size (e.g., "Standard_B16s_v2")
image_publisher	str	Yes	OS image publisher (e.g., "canonical")
image_offer	str	Yes	OS image offer (e.g., "ubuntu-24_04-lts")
image_sku	str	Yes	OS image SKU (e.g., "server")
image_version	str	Yes	OS image version (e.g., "latest")
disk_type	str	Yes	Disk storage type (e.g., "Standard_LRS")
max_undead	int	Yes	Maximum unprocessed VMs during initialization
delete_on_complete	bool	Yes	Whether VMs self-delete after trial completion
batch_id	int	Yes	Random batch identifier for naming VMs and NICs

Outputs

Name	Type	Description
assign_contributor_permissions return	Tuple	Updated (compute_client, auth_client) tuple
run_azure_process return	None	Function completes when all VMs are provisioned and permissions assigned

Usage Examples

# This module is invoked internally by AzureVMInstance executor.
# Direct usage is not recommended. See AzureVMInstance for the public API.

from powerlift.executors.azure_vm import AzureVMInstance

executor = AzureVMInstance(
    store=store,
    azure_tenant_id="...",
    subscription_id="...",
    azure_client_id="...",
    azure_client_secret="...",
    n_instances=2,
    vm_size="Standard_NC24ads_A100_v4",
)
# submit() internally calls run_azure_process via multiprocessing
executor.submit(experiment_id=1, timeout=7200)
executor.join()

Related Pages

• Interpretml_Interpret_Powerlift_AzureVMInstance -- Executor class that invokes this module
• Interpretml_Interpret_Powerlift_RunAzureCI -- Parallel implementation for Azure Container Instance provisioning
• Interpretml_Interpret_Powerlift_RunTrials -- Trial execution logic invoked inside the VMs
• Interpretml_Interpret_Powerlift_Schema -- Database models accessed by the startup script via psql