Databricks on Azure with Terraform

Brendan Thompson • 4 August 2021 • 15 min read

Overview#

This post aims to provide a walk-through of how to deploy a Databricks cluster on Azure with its supporting infrastructure using Terraform. At the end of this post, you will have all the components required to be able to complete the Tutorial: Extract, transform, and load data by using Azure Databricks tutorial on the Microsoft website.

Once we have built out the infrastructure I will lay out the steps and results that the aforementioned tutorial goes through.

I will go through building out the Terraform code file by file, below are the files you will end up with:

.
├── aad.tf
├── databricks.tf
├── main.tf
├── network.tf
├── outputs.tf
├── providers.tf
├── synapse.tf
├── variables.tf
└── versions.tf

The below diagram outlines the high-level components that will be deployed in this post.

Providers#

A key file here is providers.tf, this is where we will declare the providers we require and any configuration parameters that we need to pass to them.

providers.tf
provider "azurerm" {
  features {}
}

provider "databricks" {
  azure_workspace_resource_id = azurerm_databricks_workspace.this.id
}

As can be seen here we are setting the azurerm providers features attribute to be an empty object, and telling databricks where to find the ID for the azurerm_databricks_workspace resource.

Versions#

Another pretty important file in modern Terraform is versions.tf this is where we tell Terraform what version of Terraform it must use as well constraints we may have about any providers. For example, we may want to specify a provider version, or we may be hosting a version/fork of the provider somewhere other than the default Terraform registry.

versions.tf
terraform {
  required_version = ">= 1.0"

  required_providers {
    azurerm = {
      source = "hashicorp/azurerm"
    }
    azuread = {
      source  = "hashicorp/azuread"
      version = "~> 1.0"
    }
    databricks = {
      source = "databrickslabs/databricks"
    }
  }
}

Variables#

Now that we have our providers and any constraints setup about our Terraform environment, it is time to move onto what information we pass into our code, this is done by variables.tf.

variables.tf
variable "location" {
  type        = string
  description = "(Optional) The location for resource deployment"
  default     = "australiaeast"
}

variable "environment" {
  type        = string
  description = "(Required) Three character environment name"

  validation {
    condition     = length(var.environment) <= 3
    error_message = "Err: Environment cannot be longer than three characters."
  }
}

variable "project" {
  type        = string
  description = "(Required) The project name"
}

variable "databricks_sku" {
  type        = string
  description = <<EOT
    (Optional) The SKU to use for the databricks instance"

    Default: standard
EOT

  validation {
    condition     = can(regex("standard|premium|trial", var.databricks_sku))
    error_message = "Err: Valid options are 'standard', 'premium' or 'trial'."
  }
}

Above we have declared four variables, some of which are required and some that are optional. I find it good practice to mark a variable as optional or required in the description. On two of the variables validation has been added to ensure that what is being passed into our code is what Azure expects it to be. Doing this helps surface errors earlier.

Main#

main.tf is where anything that is foundational will usually live, an example of this for Azure is broadly an azurerm_resource_group as this is likely to be consumed by any other code written. By splitting out the code into different files it helps developers more easily understand what is going on in the codebase.

In our main.tf as can be seen below we are declaring a few things.

main.tf
locals {
  naming = {
    location = {
      "australiaeast" = "aue"
    }
  }
}

data "azurerm_client_config" "current" {}

resource "azurerm_resource_group" "this" {
  name = format("rg-%s-%s-%s",
  local.naming.location[var.location], var.environment, var.project)

  location = var.location
}

resource "azurerm_key_vault" "this" {
  name = format("kv-%s-%s-%s",
  local.naming.location[var.location], var.environment, var.project)

  resource_group_name = azurerm_resource_group.this.name
  location            = azurerm_resource_group.this.location
  tenant_id           = data.azurerm_client_config.current.tenant_id

  sku_name = "standard"

  access_policy {
    tenant_id = data.azurerm_client_config.current.tenant_id
    object_id = data.azurerm_client_config.current.object_id

    secret_permissions = [
      "Get",
      "Set",
      "Delete",
      "Recover",
      "Purge"
    ]
  }
}

resource "azurerm_databricks_workspace" "this" {
  name = format("dbs-%s-%s-%s",
  local.naming.location[var.location], var.environment, var.project)

  resource_group_name = azurerm_resource_group.this.name
  location            = azurerm_resource_group.this.location
  sku                 = var.databricks_sku

  custom_parameters {
    virtual_network_id  = azurerm_virtual_network.this.id
    public_subnet_name  = azurerm_subnet.public.name
    private_subnet_name = azurerm_subnet.private.name
  }

  depends_on = [
    azurerm_subnet_network_security_group_association.public,
    azurerm_subnet_network_security_group_association.private,
  ]
}

As can be seen above we use the following format() blocks to more consistently name our resources. This will help give engineers more confidence in the naming of resources, and if there is standard that is kept to across the platform it will enable them to more easily orientate themselves within any environment on the platform.

format("<Resource_Identifier>-%s-%s-%s",
  local.naming.location[var.location], var.environment, var.project)

Network#

Now that we have the code ready for our Databricks workspace we need to create the network as you can see that we are referencing those in our main.tf file. The types of resources that we are creating and their purpose are as follows;

network.tf
resource "azurerm_virtual_network" "this" {
  name = format("vn-%s-%s-%s",
  local.naming.location[var.location], var.environment, var.project)

  location            = azurerm_resource_group.this.location
  resource_group_name = azurerm_resource_group.this.name

  address_space = ["10.0.0.0/16"]
}

resource "azurerm_subnet" "private" {
  name = format("sn-%s-%s-%s-priv",
  local.naming.location[var.location], var.environment, var.project)

  resource_group_name  = azurerm_resource_group.this.name
  virtual_network_name = azurerm_virtual_network.this.name
  address_prefixes     = ["10.0.0.0/24"]

  delegation {
    name = "databricks-delegation"

    service_delegation {
      name = "Microsoft.Databricks/workspaces"
      actions = [
        "Microsoft.Network/virtualNetworks/subnets/join/action",
        "Microsoft.Network/virtualNetworks/subnets/prepareNetworkPolicies/action",
        "Microsoft.Network/virtualNetworks/subnets/unprepareNetworkPolicies/action",
      ]
    }
  }
}

resource "azurerm_network_security_group" "private" {
  name = format("nsg-%s-%s-%s-priv",
  local.naming.location[var.location], var.environment, var.project)

  resource_group_name = azurerm_resource_group.this.name
  location            = azurerm_resource_group.this.location
}

resource "azurerm_subnet_network_security_group_association" "private" {
  subnet_id                 = azurerm_subnet.private.id
  network_security_group_id = azurerm_network_security_group.private.id
}

resource "azurerm_subnet" "public" {
  name = format("sn-%s-%s-%s-pub",
  local.naming.location[var.location], var.environment, var.project)

  resource_group_name  = azurerm_resource_group.this.name
  virtual_network_name = azurerm_virtual_network.this.name
  address_prefixes     = ["10.0.1.0/24"]

  delegation {
    name = "databricks-delegation"

    service_delegation {
      name = "Microsoft.Databricks/workspaces"
      actions = [
        "Microsoft.Network/virtualNetworks/subnets/join/action",
        "Microsoft.Network/virtualNetworks/subnets/prepareNetworkPolicies/action",
        "Microsoft.Network/virtualNetworks/subnets/unprepareNetworkPolicies/action",
      ]
    }
  }
}

resource "azurerm_network_security_group" "public" {
  name = format("nsg-%s-%s-%s-pub",
  local.naming.location[var.location], var.environment, var.project)

  resource_group_name = azurerm_resource_group.this.name
  location            = azurerm_resource_group.this.location
}

resource "azurerm_subnet_network_security_group_association" "public" {
  subnet_id                 = azurerm_subnet.public.id
  network_security_group_id = azurerm_network_security_group.public.id
}

The network sizes declared above are rather large and ideally should not be used in any production environment, however, they are perfectly fine for development or proof of concept work as long as the network is never peered or connected to an express route where it might conflict with internal ranges.

Azure Active Directory (AAD)#

Before we look into creating the internals of our Databricks instance or our Azure Synapse database we must first create the Azure Active Directory (AAD) application that will be used for authentication.

aad.tf
resource "random_password" "service_principal" {
  length  = 16
  special = true
}

resource "azurerm_key_vault_secret" "service_principal" {
  name         = "service-principal-password"
  value        = random_password.service_principal.result
  key_vault_id = azurerm_key_vault.this.id
}

resource "azuread_application" "this" {
  display_name = format("app-%s-%s-%s",
  local.naming.location[var.location], var.environment, var.project)
}

resource "azuread_service_principal" "this" {
  application_id               = azuread_application.this.application_id
  app_role_assignment_required = false
}

resource "azuread_service_principal_password" "this" {
  service_principal_id = azuread_service_principal.this.id
  value                = azurerm_key_vault_secret.service_principal.value
}

resource "azurerm_role_assignment" "this" {
  scope                = azurerm_storage_account.this.id
  role_definition_name = "Storage Blob Data Contributor"
  principal_id         = azuread_service_principal.this.object_id
}

Above we are creating resources with the following properties:

Synapse#

Now that we have our credentials all ready to go we can setup the Synapse instance, as well as any ancillary resources that we might need.

synapse.tf
resource "azurerm_storage_account" "this" {
  name = format("sa%s%s%s",
  local.naming.location[var.location], var.environment, var.project)

  resource_group_name = azurerm_resource_group.this.name
  location            = azurerm_resource_group.this.location

  account_tier             = "Standard"
  account_replication_type = "LRS"
  account_kind             = "BlobStorage"
}

resource "azurerm_storage_data_lake_gen2_filesystem" "this" {
  name = format("fs%s%s%s",
  local.naming.location[var.location], var.environment, var.project)
  storage_account_id = azurerm_storage_account.this.id
}

resource "azurerm_key_vault_secret" "sql_administrator_login" {
  name         = "sql-administrator-login"
  value        = "sqladmin"
  key_vault_id = azurerm_key_vault.this.id
}

resource "random_password" "sql_administrator_login" {
  length  = 16
  special = false
}

resource "azurerm_key_vault_secret" "sql_administrator_login_password" {
  name         = "sql-administrator-login-password"
  value        = random_password.sql_administrator_login.result
  key_vault_id = azurerm_key_vault.this.id
}

resource "azurerm_synapse_workspace" "this" {
  name = format("ws-%s-%s-%s",
  local.naming.location[var.location], var.environment, var.project)

  resource_group_name                  = azurerm_resource_group.this.name
  location                             = azurerm_resource_group.this.location
  storage_data_lake_gen2_filesystem_id = azurerm_storage_data_lake_gen2_filesystem.this.id

  aad_admin = [
    {
      login     = "AzureAD Admin"
      object_id = azuread_service_principal.this.object_id
      tenant_id = data.azurerm_client_config.current.tenant_id
    }
  ]

  sql_administrator_login          = azurerm_key_vault_secret.sql_administrator_login.value
  sql_administrator_login_password = azurerm_key_vault_secret.sql_administrator_login_password.value
}

resource "azurerm_synapse_sql_pool" "this" {
  name = format("pool_%s",
  var.project)

  synapse_workspace_id = azurerm_synapse_workspace.this.id
  sku_name             = "DW100c"
  create_mode          = "Default"
}

resource "azurerm_synapse_firewall_rule" "allow_azure_services" {
  name                 = "AllowAllWindowsAzureIps"
  synapse_workspace_id = azurerm_synapse_workspace.this.id
  start_ip_address     = "0.0.0.0"
  end_ip_address       = "0.0.0.0"
}

As per usual we will go through each of the resources being created and explain what they do.

Databricks#

Finally from a resource creation perspective we need to setup the internals of the Databricks instance. This mostly entails creating a single node Databricks cluster where Notebooks etc can be created by Data Engineers.

databricks.tf
data "databricks_node_type" "smallest" {
  local_disk = true

  depends_on = [
    azurerm_databricks_workspace.this
  ]
}

data "databricks_spark_version" "latest_lts" {
  long_term_support = true

  depends_on = [
    azurerm_databricks_workspace.this
  ]
}

resource "databricks_cluster" "this" {
  cluster_name = format("dbsc-%s-%s-%s",
  local.naming.location[var.location], var.environment, var.project)

  spark_version = data.databricks_spark_version.latest_lts.id
  node_type_id  = data.databricks_node_type.smallest.id

  autotermination_minutes = 20

  spark_conf = {
    "spark.databricks.cluster.profile" : "singleNode"
    "spark.master" : "local[*]"
  }

  custom_tags = {
    "ResourceClass" = "SingleNode"
  }
}

Outputs#

The last thing we will need to write in Terraform will be our outputs.tf, this is the information we want returned to us once the deployment of all the previous code is complete.

outputs.tf
output "azure_details" {
  sensitive = true
  value = {
    tenant_id     = data.azurerm_client_config.current.tenant_id
    client_id     = azuread_application.this.application_id
    client_secret = azurerm_key_vault_secret.service_principal.value
  }
}

output "storage_account" {
  sensitive = true
  value = {
    name           = azurerm_storage_account.this.name
    container_name = azurerm_storage_data_lake_gen2_filesystem.this.name
    access_key     = azurerm_storage_account.this.primary_access_key
  }
}

output "synapse" {
  sensitive = true
  value = {
    database = azurerm_synapse_sql_pool.this.name
    host     = azurerm_synapse_workspace.this.connectivity_endpoints
    user     = azurerm_synapse_workspace.this.sql_administrator_login
    password = azurerm_synapse_workspace.this.sql_administrator_login_password
  }
}

In this we are simply outputting information such as our Service Principal details, information about our storage account and Synapse instance and how to authenticate to them.

Deploying the environment#

Now that we have all the pieces ready for us to use we can deploy it. This assumes that the files are all in your local directory and that you have Terraform installed.

  1. Firstly we will need to initialize terraform and pull down all the providers
terraform init
  1. Plan the deployment
terraform plan -var="environment=dev" -var="project=meow"
  1. Apply the deployment
terraform apply -var="environment=dev" -var="project=meow"

Running an ETL in Databricks#

Now that we have our environment deployed we can run through the ETL tutorial from Microsoft I linked at the top of this page.

  1. From the Azure portal within the Databricks resource click on Launch Workspace

    Launch Databricks workspace
  2. On the Databricks summary page click on New notebook

    Create Databricks notebook
  3. On the open dialogue give the notebook a name, select Scala and then select the cluster we just created

  4. From within the notebook in the first Cell but in the following code which will setup the session configuration

val appID = "<appID>"
val secret = "<secret>"
val tenantID = "<tenant-id>"

spark.conf.set("fs.azure.account.auth.type", "OAuth")
spark.conf.set("fs.azure.account.oauth.provider.type", "org.apache.hadoop.fs.azurebfs.oauth2.ClientCredsTokenProvider")
spark.conf.set("fs.azure.account.oauth2.client.id", "<appID>")
spark.conf.set("fs.azure.account.oauth2.client.secret", "<secret>")
spark.conf.set("fs.azure.account.oauth2.client.endpoint", "https://login.microsoftonline.com/<tenant-id>/oauth2/token")
spark.conf.set("fs.azure.createRemoteFileSystemDuringInitialization", "true")

The result of the above command is below, and it shows the configuration has been written.

  1. In the next cell we will setup the account configuration to allow connectivity to our storage accounts
val storageAccountName = "<storage-account-name>"
val appID = "<app-id>"
val secret = "<secret>"
val fileSystemName = "<file-system-name>"
val tenantID = "<tenant-id>"

spark.conf.set("fs.azure.account.auth.type." + storageAccountName + ".dfs.core.windows.net", "OAuth")
spark.conf.set("fs.azure.account.oauth.provider.type." + storageAccountName + ".dfs.core.windows.net", "org.apache.hadoop.fs.azurebfs.oauth2.ClientCredsTokenProvider")
spark.conf.set("fs.azure.account.oauth2.client.id." + storageAccountName + ".dfs.core.windows.net", "" + appID + "")
spark.conf.set("fs.azure.account.oauth2.client.secret." + storageAccountName + ".dfs.core.windows.net", "" + secret + "")
spark.conf.set("fs.azure.account.oauth2.client.endpoint." + storageAccountName + ".dfs.core.windows.net", "https://login.microsoftonline.com/" + tenantID + "/oauth2/token")
spark.conf.set("fs.azure.createRemoteFileSystemDuringInitialization", "true")
dbutils.fs.ls("abfss://" + fileSystemName  + "@" + storageAccountName + ".dfs.core.windows.net/")
spark.conf.set("fs.azure.createRemoteFileSystemDuringInitialization", "false")

We can see below that the account configuration has also been written.

  1. Now the sample data needs to be downloaded onto the cluster
%sh wget -P /tmp https://raw.githubusercontent.com/Azure/usql/master/Examples/Samples/Data/json/radiowebsite/small_radio_json.json

It can be seen below that the file has been downloaded to temporary storage on the cluster.

  1. Now we need to transfer from the cluster to the Azure Data Lake Storage (ADLS)
dbutils.fs.cp("file:///tmp/small_radio_json.json", "abfss://" + fileSystemName + "@" + storageAccountName + ".dfs.core.windows.net/")

The result of this command can be seen below as true.

  1. Extract the data from ADLS into a dataframe
val df = spark.read.json("abfss://" + fileSystemName + "@" + storageAccountName + ".dfs.core.windows.net/small_radio_json.json")

From the result of the above command we can see that the data is now in a dataframe.

  1. Show the contents of the dataframe
df.show()

Below it can be seen that the dataframe contains the data from the sample file.

  1. The first basic transformation we will do is to only retrieve specific columns from the originally extracted dataframe
val specificColumnsDf = df.select("firstname", "lastname", "gender", "location", "level")
specificColumnsDf.show()

You can see below that the only columns that are now returned in the dataframe are:

  1. The second transformation that we will do is to rename some columns
val renamedColumnsDF = specificColumnsDf.withColumnRenamed("level", "subscription_type")
renamedColumnsDF.show()

Below shows that the second transformation we have performed is to rename the column level to subscription_type.

  1. Now we are going to load that transformed dataframe into Azure Synapse for later use
// Storage Account
val blobStorage = "<blob-storage-account-name>.blob.core.windows.net"
val blobContainer = "<blob-container-name>"
val blobAccessKey =  "<access-key>"

val tempDir = "wasbs://" + blobContainer + "@" + blobStorage +"/tempDirs"

val acntInfo = "fs.azure.account.key."+ blobStorage
sc.hadoopConfiguration.set(acntInfo, blobAccessKey)

// Azure Synapse
val dwDatabase = "<database-name>"
val dwServer = "<database-server-name>"
val dwUser = "<user-name>"
val dwPass = "<password>"
val dwJdbcPort =  "1433"
val dwJdbcExtraOptions = "encrypt=true;trustServerCertificate=true;hostNameInCertificate=*.database.windows.net;loginTimeout=30;"
val sqlDwUrl = "jdbc:sqlserver://" + dwServer + ":" + dwJdbcPort + ";database=" + dwDatabase + ";user=" + dwUser+";password=" + dwPass + ";$dwJdbcExtraOptions"
val sqlDwUrlSmall = "jdbc:sqlserver://" + dwServer + ":" + dwJdbcPort + ";database=" + dwDatabase + ";user=" + dwUser+";password=" + dwPass

spark.conf.set(
    "spark.sql.parquet.writeLegacyFormat",
    "true")

renamedColumnsDF.write.format("com.databricks.spark.sqldw")
.option("url", sqlDwUrlSmall)
.option("dbtable", "SampleTable")
.option( "forward_spark_azure_storage_credentials","True")
.option("tempdir", tempDir)
.mode("overwrite")
.save()

This last screenshot shows that the data has been pushed over to our Azure Synapse instance.

Brendan Thompson

Principal Cloud Engineer

Discuss on Twitter