MLOps Azure Infrastructure Module with Service Principal Linking

This module sets up multi-workspace model registry between a development (dev) workspace, a staging workspace, and a production (prod) workspace, allowing READ access from dev/staging workspaces to staging & prod model registries.

The module performs this setup by linking pre-existing AAD applications with newly created Azure Databricks service principals in the staging and prod workspaces, then giving them READ-only access to their respective model registries. It will also create secret scopes and store the necessary secrets in the dev and staging workspaces, and only give READ access to this secret scope to the "users" group and the generated service principals group. The output of this module will be the secret scope names and prefixes since these values are needed to be able to access the remote model registry.

NOTE:

1. This module is in preview so it is still experimental and subject to change. Feedback is welcome!
2. The Databricks providers that are passed into the module must be configured with workspace admin permissions.
3. In order to create tokens for service principals, they are added to a group, which is then given token_usage permission. However, in order to set this permission, there must be at least 1 personal access token in the workspace, and this permission strictly overwrites existing permissions. Currently, running this module will overwrite permissions to allow token usage only for members of the generated service principals group in the staging and prod workspaces. If additional groups are desired to have token_usage permissions, they can be set via the additional_token_usage_groups input variable.
4. The service principal tokens stored for remote model registry access are created with a default expiration of 100 days (8640000 seconds), and the module will need to be re-applied after this time to refresh the tokens.

Usage

Option 1: Authentication with Azure Client Secret and Tenant ID

This option will use the client secrets and tenant IDs to generate AAD tokens for authentication. The advantage to this approach is not having to manually generate short-lived AAD tokens (normally maximum lifetime of 60 minutes) each time this module needs to be used.

NOTE: This option requires that Python 3.8+ be installed to obtain the service principal's AAD token.

provider "databricks" {
  alias = "dev"    # Authenticate using preferred method as described in Databricks provider
}

provider "databricks" {
  alias = "staging"     # Authenticate using preferred method as described in Databricks provider
}

provider "databricks" {
  alias = "prod"     # Authenticate using preferred method as described in Databricks provider
}

module "mlops_azure_infrastructure_with_sp_linking" {
  source = "databricks/mlops-azure-infrastructure-with-sp-linking/databricks"
  providers = {
    databricks.dev = databricks.dev
    databricks.staging = databricks.staging
    databricks.prod = databricks.prod
  }
  staging_workspace_id          = "123456789"
  prod_workspace_id             = "987654321"
  azure_staging_client_id       = "k9l8m7n6o5-e5f6-g7h8-i9j0-a1b2c3d4p4"
  azure_staging_client_secret   = var.azure_staging_client_secret     # This value is sensitive.
  azure_staging_tenant_id       = "a1b2c3d4-e5f6-g7h8-i9j0-k9l8m7n6o5p4"
  azure_prod_client_id          = "k9l8m7n6p4-e5f6-g7h8-i9j0-a1b2c3d4o5"
  azure_prod_client_secret      = var.azure_prod_client_secret     # This value is sensitive.
  azure_prod_tenant_id          = "a1b2c3d4-e5f6-g7h8-i9j0-k9l8m7n6o5p4"
  additional_token_usage_groups = ["users"]     # This field is optional.
}

Option 2: Authentication with Azure Active Directory (AAD) Token

This option will use the provided AAD tokens for authentication. The advantage to this approach is not having to create/provide client secrets and tenant IDs.

provider "databricks" {
  alias = "dev"    # Authenticate using preferred method as described in Databricks provider
}

provider "databricks" {
  alias = "staging"     # Authenticate using preferred method as described in Databricks provider
}

provider "databricks" {
  alias = "prod"     # Authenticate using preferred method as described in Databricks provider
}

module "mlops_azure_infrastructure_with_sp_linking" {
  source = "databricks/mlops-azure-infrastructure-with-sp-linking/databricks"
  providers = {
    databricks.dev = databricks.dev
    databricks.staging = databricks.staging
    databricks.prod = databricks.prod
  }
  staging_workspace_id          = "123456789"
  prod_workspace_id             = "987654321"
  azure_staging_client_id       = "k9l8m7n6o5-e5f6-g7h8-i9j0-a1b2c3d4p4"
  azure_staging_aad_token       = var.azure_staging_aad_token     # This value is sensitive.
  azure_prod_client_id          = "k9l8m7n6p4-e5f6-g7h8-i9j0-a1b2c3d4o5"
  azure_prod_aad_token          = var.azure_prod_aad_token     # This value is sensitive.
  additional_token_usage_groups = ["users"]     # This field is optional.
}

Usage example with MLOps Azure Project Module with Service Principal Linking

provider "databricks" {
  alias = "dev"    # Authenticate using preferred method as described in Databricks provider
}

provider "databricks" {
  alias = "staging"     # Authenticate using preferred method as described in Databricks provider
}

provider "databricks" {
  alias = "prod"     # Authenticate using preferred method as described in Databricks provider
}

module "mlops_azure_infrastructure_with_sp_linking" {
  source = "databricks/mlops-azure-infrastructure-with-sp-linking/databricks"
  providers = {
    databricks.dev = databricks.dev
    databricks.staging = databricks.staging
    databricks.prod = databricks.prod
  }
  staging_workspace_id          = "123456789"
  prod_workspace_id             = "987654321"
  azure_staging_client_id       = "k9l8m7n6o5-e5f6-g7h8-i9j0-a1b2c3d4p4"
  azure_staging_client_secret   = var.azure_staging_client_secret     # This value is sensitive.
  azure_staging_tenant_id       = "a1b2c3d4-e5f6-g7h8-i9j0-k9l8m7n6o5p4"
  azure_prod_client_id          = "k9l8m7n6p4-e5f6-g7h8-i9j0-a1b2c3d4o5"
  azure_prod_client_secret      = var.azure_prod_client_secret     # This value is sensitive.
  azure_prod_tenant_id          = "a1b2c3d4-e5f6-g7h8-i9j0-k9l8m7n6o5p4"
  additional_token_usage_groups = ["users"]     # This field is optional.
}

module "mlops_azure_project_with_sp_linking" {
  source = "databricks/mlops-azure-project-with-sp-linking/databricks"
  providers = {
    databricks.staging = databricks.staging
    databricks.prod = databricks.prod
  }
  service_principal_name      = "example-name"
  project_directory_path      = "/dir-name"
  azure_staging_client_id     = "k9l8m7n6o5-e5f6-g7h8-i9j0-a1b2c3d4p4"
  azure_staging_client_secret = var.azure_staging_client_secret     # This value is sensitive.
  azure_staging_tenant_id     = "a1b2c3d4-e5f6-g7h8-i9j0-k9l8m7n6o5p4"
  azure_prod_client_id        = "k9l8m7n6p4-e5f6-g7h8-i9j0-a1b2c3d4o5"
  azure_prod_client_secret    = var.azure_prod_client_secret     # This value is sensitive.
  azure_prod_tenant_id        = "a1b2c3d4-e5f6-g7h8-i9j0-k9l8m7n6o5p4"
  service_principal_group_name = module.mlops_azure_infrastructure_with_sp_linking.service_principal_group_name 
  # The above field is optional, especially since in this case service_principal_group_name will be mlops-service-principals either way, 
  # but this also serves to create an implicit dependency. Can also be replaced with the following line to create an explicit dependency:
  # depends_on             = [module.mlops_azure_infrastructure_with_sp_linking]
}

Requirements

Name	Version
terraform	>=1.1.6
databricks	>=0.5.8
python <Option 1>	>=3.8

Inputs

Name	Description	Type	Default	Required
staging_workspace_id	Workspace ID of the staging workspace (can be often found in the URL) used for remote model registry setup.	string	N/A	yes
prod_workspace_id	Workspace ID of the prod workspace (can be often found in the URL) used for remote model registry setup.	string	N/A	yes
azure_staging_client_id	The client ID of the AAD service principal in the staging workspace that will be used to READ the model registry from the dev workspace.	string	N/A	yes
azure_staging_aad_token	The AAD token of the service principal in the staging workspace. This will need to be manually refreshed once it expires (often within several minutes). NOTE: Only this or both azure_staging_client_secret & azure_staging_tenant_id need to be provided.	string	null	no
azure_staging_client_secret	The client secret of the AAD service principal in the staging workspace. NOTE: If azure_staging_aad_token is not provided, this and azure_staging_tenant_id must be provided to generate an AAD token.	string	null	no
azure_staging_tenant_id	The tenant ID of the AAD service principal in the staging workspace. NOTE: If azure_staging_aad_token is not provided, this and azure_staging_client_secret must be provided to generate an AAD token.	string	null	no
azure_prod_client_id	The client ID of the AAD service principal in the prod workspace that will be used to READ the model registry from the dev & staging workspaces.	string	N/A	yes
azure_prod_aad_token	The AAD token of the service principal in the prod workspace. This will need to be manually refreshed once it expires (often within several minutes). NOTE: Only this or both azure_prod_client_secret & azure_prod_tenant_id need to be provided.	string	null	no
azure_prod_client_secret	The client secret of the AAD service principal in the prod workspace. NOTE: If azure_prod_aad_token is not provided, this and azure_prod_tenant_id must be provided to generate an AAD token.	string	null	no
azure_prod_tenant_id	The tenant ID of the AAD service principal in the prod workspace. NOTE: If azure_prod_aad_token is not provided, this and azure_prod_client_secret must be provided to generate an AAD token.	string	null	no
additional_token_usage_groups	List of groups that should have token usage permissions in the staging and prod workspaces, along with the created service principal group (mlops-service-principals).	list(string)	[]	no

Outputs

Name	Description	Type	Sensitive
dev_secret_scope_name_for_staging	The name of the secret scope created in the dev workspace that is used for remote model registry access to the staging workspace.	string	no
dev_secret_scope_name_for_prod	The name of the secret scope created in the dev workspace that is used for remote model registry access to the prod workspace.	string	no
staging_secret_scope_name_for_prod	The name of the secret scope created in the staging workspace that is used for remote model registry access to the prod workspace.	string	no
dev_secret_scope_prefix_for_staging	The prefix used in the dev workspace secret scope for remote model registry access to the staging workspace.	string	no
dev_secret_scope_prefix_for_prod	The prefix used in the dev workspace secret scope for remote model registry access to the prod workspace.	string	no
staging_secret_scope_prefix_for_prod	The prefix used in the staging workspace secret scope for remote model registry access to the prod workspace.	string	no
service_principal_group_name	The name of the service principal group created in the staging and prod workspace.	string	no

Providers

Name	Authentication	Use
databricks.dev	Provided by the user.	Generate all resources in the dev workspace.
databricks.staging	Provided by the user.	Generate all resources in the staging workspace.
databricks.prod	Provided by the user.	Generate all resources in the prod workspace.
databricks.staging_sp	Authenticated via host and generated AAD token for service principal.	Obtain service principal PAT.
databricks.prod_sp	Authenticated via host and generated AAD token for service principal.	Obtain service principal PAT.

Resources

Name	Type
databricks_current_user.staging_user	data source
databricks_current_user.prod_user	data source
databricks_group.staging_sp_group	resource
databricks_permissions.staging_token_usage	resource
databricks_group.prod_sp_group	resource
databricks_permissions.prod_token_usage	resource
databricks_token.staging_sp_token	resource
databricks_token.prod_sp_token	resource
azure-link-service-principal.link_staging_sp	module
azure-link-service-principal.link_prod_sp	module
remote-model-registry.remote_model_registry_dev_to_staging	module
remote-model-registry.remote_model_registry_dev_to_prod	module
remote-model-registry.remote_model_registry_staging_to_prod	module

Known Issues

• AAD token generation occasionally fails with "HTTP Error 400: Bad Request" but the query is not actually invalid.
  • Solution: Re-run terraform apply and the error should disappear.