Provisioning Azure Databricks workspace with Terraform
Terraform provides an
azurerm_databricks_workspace resource to create an Azure Databricks
The workspace itself is simple enough, but there are a number of other related resources you may need to provision with Terraform to get a fully working environment.
Private endpoints to storage
If you use VNet injection (also known as “no public IP”) with Databricks, you will probably want Databricks to use a private endpoint to access your storage account, to ensure a direct path from Databricks to your storage regardless of VNet egress configuration.
This can be provisioned through Terraform as well with the following resources:
Note that there is nothing special about Databricks here, other than that you will need a separate subnet for the private endpoint, in the same VNet. You can’t create the private endpoint in either of the two subnets you create with delegation to Databricks.
The other note is that, because storage accounts only allow one subresource name
(despite the argument being a list in Terraform) you would need a separate endpoint if you
dfs to work.
Once this is set up you can validate that Databricks is indeed hitting your private endpoint by running a shell command in a notebook like
%sh ping mystorageaccount.dfs.core.windows.net
And see that the IP address is an IP in your private endpoint subnet.
Another best-practice is to use User-Defined Routes (UDRs) with your Databricks VNet to provide the most direct path to the Databricks control plane. Otherwise there’s a risk you may inadvertently “trombone” through your on-prem network if you have VNet peering to an Express Route or VPN.
UDRs are not always strictly required, but if your VNet does not account for egress to the control plane somehow, you may see timeout issues trying to launch new Databricks clusters. At the very least, your VNet configuration needs to account for egress to the control plane somehow.
Databricks Unity is the ideal way to connect to ADLS2 with managed identity. If this isn’t an option, you can connect Databricks to ADLS2 with a service principal and secret. The secrets can be stored in Azure Key Vault and Databricks can reference the Key Vault secrets with a secret scope.
First, create Terraform resources to create the service principal and secret:
azuread_service_principal(references application above)
azuread_application_password(references application above)
azurerm_key_vault_secret(references secret and puts it in Key Vault)
Then grant the service principal access to the storage account:
azurerm_role_assignmentwith “Storage Blob Data Reader/Contributor” (references storage account and service principal)
Sidebar: The distinction between “application” and “service principal” in Azure can be confusing – the concepts are linked but different. The Application is the resource with an OAuth client ID and secret and the service principal is what is granted access to other Azure resources.
Once you have the Azure resources you can create a secret scope in Databricks.
Previously it was not possible to do this with the Databricks provider resource
databricks_secret_scope (“Currently, it’s only possible to create Azure Key Vault scopes with Azure CLI authentication and not with Service Principal…”)
This may have changed recently.
However, using the
databricks Terraform provider needs to be done with caution,
as you can run into issues if you have direct dependencies between
azurerm_databricks_workspace resources and the
This is a general challenge with Terraform not limited to Databricks.