Experimenting with Terramate and SAP BTP

UPDATE 13.12.2024

I updated several parts of this blog post as I missed one feature of Terramate concerning the output sharing namely the option to mock input. With this feature in place and the corresponding setting in the scripts, the challenge of setting up the stacks with dynamic dependencies worked.

The Challenge

In case you are working with Terraform in a multi-provider setup you certainly also came across the following challenge:

You want to create a resource 1 that depends on provider A and a resource 2 that depends on provider B. So far, so good, but the read-only data of resource 1 contains parts of the information that is needed to configure provider B. Resource 2 has a dependency to resource 1, so from a logical perspective things should work out, as from a runtime perspective provider B is needed once resource 1 and consequently all relevant information is available.

Unfortunately, Terraform as well as OpenTofu require the provider configuration fully in place from the start. A dynamic configuration or a lazy loading of the provider configuration is not possible, and I did not come across any plans to provide such a functionality.

In this blog post I want to present a possible solution leveraging Terramate. As an example, I will be using one example from my "daily life" namely SAP BTP and the Cloud Foundry environment on SAP BTP. This setup perfectly matches the above challenge:

• The Cloud Foundry environment (namely the organization) can be created via the Terraform provider for SAP BTP. After the creation, we have access to the data relevant for authenticating against the Cloud Foundry provider namely the API endpoint of the environment.
• In addition, we need another information namely the ID of the Cloud Foundry organization to setup further resources in Cloud Foundry.

So, jackpot ... this is a challenge and it is for real. Let's see how we can resolve it or at least make life a bit easier.

Status Quo

Currently if we want to resolve the situation, we must execute Terraform in two steps:

1. Setup the Cloud Foundry environment via the Terraform Provider for SAP BTP.
2. Do the consequent setup in Cloud Foundry via a second Configuration, transferring the information from the first step to the second one e.g., via outputs.

While this is doable in CI/CD pipelines e.g. via GitHub Actions, testing things locally gets painful. As a workaround you can write the output of step 1 to a terraform.tfvars file of step 2, which gives some level of convenience, but with this setup the setup in the CI/CD scenario differs from your local setup which is also not desirable.

This is where I think Terramate can help. Although it cannot do some black magic and make things work as we would love them to, we can get a more streamlined setup. What do we need for that? Let's first explore some (experimental) features of Terramate that can help us.

Terramate (Experimental) Features

Terramate as a productivity tool on top of Terraform (or OpenTofu) can support us with the orchestration of the Terraform flow. Our recipe we want to use contains the following ingredients:

• Stacks which we will use to define self-contained units for the deployment
• Explicit Order of Execution which we use to ensure that in case of a full deployment the self-contained units are executed in the right sequence
• Outputs Sharing(experimental) for transferring the output of one stack to the dependent one
• Scripts(experimental) to split the Terraform specific commands from the Terramate commands and to avoid multiple commands to be keyed in reducing the chance to make a mistake.

As you can see we will leverage some experimental features which "might still be subject to changes in the future.". We love to live on the edge of new features, so we accept that, right?

Let's see how we can bring these ingredients together.

Terramating the Experience of SAP BTP and Cloud Foundry

Let us shortly recap the setup that we plan to have:

• On the SAP BTP side of the house we first want to setup a subaccount and an envrionment of type Cloud Foundry. This is done via the Terraform provider for SAP BTP.
• After that we want to create a space in the Cloud Foundry environment and if applicable assign some space roles. This is done via the Terraform provider for ≈Cloud Foundry that needs the output of the previous setup for the configuration.

The code samples that you will see, will focus on the Terramate features mentioned above. I did not use the code generation feature of Terramate to avoid distracting from the main topics of this blog post. Having said that, we would probably add code generation to the setup when bringing things to production.

Defining the stack

First we need to do is define the Terramate stacks. The kind of natural split is to have one stack for the subaccount and one for the Cloud Foundry setup. We achieve this via:

terramate create --name "subacccount-dev" --description "Basics for BTP development setup" --tags "subaccount,dev" stacks/subaccount_dev
terramate create --name "cloudfoundry-dev" --description "CF for BTP dev setup" --tags "cloudfoundry,dev" stacks/cloudfoundry_dev

This results in this directory structure:

| - stacks
|   | - subacount_dev
|   | - cloudfoundry_dev

Each of the stacks contains the stack information in the stack.tm.hcl file.

Excellent. Let's move on to bring them in the right execution order.

Defining the execution order

Both stacks are on the same level, so no implicit dependency and consequently no order is defined. Nevertheless, we would like to have such an order. We could either use sub-stacks or we define the order explicitly. In this blog post we make use of the second option.

We configure the order of execution via the after block in the stack definition. As we want to instruct Terramate to execute the Cloud Foundry stack after the subaccount stack, we adjust the stack.tm.hcl file accordingly:

stack {
  name        = "cloudfoundry-dev"
  description = "CF for BTP dev setup"
  tags        = ["cloudfoundry", "dev"]
  id          = "d5962b3f-3b79-412f-9970-93112741855a"
  after       = ["tag:subaccount:dev"]
}

We can specify the after block via tags as we did or via the path to the stack. Using the tags is considered as best practice and keeps us more flexible. We can validate the sequence via the command:

terramate list --run-order

With that we can add the configuration for the setup.

The Terraform configuration

I do not want to clutter this blog post with a lot of basic Terraform code, so the following snippets focus on the main bits and pieces of our storyline. You find the complete code on GitHub.

In the subaccount_dev stack we define the resources as well as the provider configuration as "usual". Although we need to define an output for the configuration to get access to the API as well as to the organization ID of the Cloud Foundry environment, we do not do that manually.

Terramate will close the gap here. For a better understanding of the following code snippets, this is the main.tf of the subaccount stack:

resource "random_uuid" "uuid" {}

locals {
  random_uuid       = random_uuid.uuid.result
  subaccount_domain = lower("${var.subaccount_name}-${local.random_uuid}")
  subaccount_name   = var.subaccount_name
  subaccount_cf_org = substr(replace("${local.subaccount_domain}", "-", ""), 0, 32)
}

resource "btp_subaccount" "sa_dev_base" {
  name      = var.subaccount_name
  subdomain = join("-", ["sa-dev-base", random_uuid.uuid.result])
  region    = lower(var.region)
}

# Fetch all available environments for the subaccount
data "btp_subaccount_environments" "all" {
  subaccount_id = btp_subaccount.sa_dev_base.id
}

# Take the landscape label from the first CF environment if no environment label is provided
resource "terraform_data" "cf_landscape_label" {
  input = [for env in data.btp_subaccount_environments.all.values : env if env.service_name == "cloudfoundry" && env.environment_type == "cloudfoundry"][0].landscape_label
}

# Create the Cloud Foundry environment instance
resource "btp_subaccount_environment_instance" "cfenv_dev_base" {
  subaccount_id    = btp_subaccount.sa_dev_base.id
  name             = local.subaccount_cf_org
  environment_type = "cloudfoundry"
  service_name     = "cloudfoundry"
  plan_name        = var.cf_plan_name
  landscape_label  = terraform_data.cf_landscape_label.output

  parameters = jsonencode({
    instance_name = local.subaccount_cf_org
  })
}

The resource of interest is the btp_subaccount_environment_instance which delivers the necessary information for the other stack.

When it comes to the cloudfoundry_dev stack, we also define the project with the usual structure, a main.tf with the space creation and the space role assignment:

resource "cloudfoundry_space" "dev_space" {
  name = var.cf_space_name
  org  = var.cf_org_id
}

resource "cloudfoundry_space_role" "space_developer" {
  for_each = toset(var.cf_space_developers)
  username = each.value
  type     = "space_developer"
  space    = cloudfoundry_space.dev_space.id
}

resource "cloudfoundry_space_role" "space_manager" {
  for_each = toset(var.cf_space_managers)
  username = each.value
  type     = "space_manager"
  space    = cloudfoundry_space.dev_space.id
}

and a provider configuration that looks like this:

terraform {
  required_providers {
    cloudfoundry = {
      source  = "sap/cloudfoundry"
      version = "1.0.0-rc1"
    }
  }
}

provider "cloudfoundry" {
  api_url = var.cf_api_url
}

We see that the API URL in the provider configuration as well as the Cloud Foundry org ID is a variable (var.cf_api_url and var.cf_org_id). Defining the variables is a bit counterintuitive now. As for the outputs in the other stack we leave these two variables out of the variables.tf file which looks like this:

variable "cf_space_name" {
  type        = string
  description = "The name of the Cloud Foundry space."
  default     = "dev"
}

variable "cf_space_managers" {
  type        = list(string)
  description = "List of managers for the Cloud Foundry space."
  default     = []
}

variable "cf_space_developers" {
  type        = list(string)
  description = "List of developers for the Cloud Foundry space."
  default     = []
}

As for the outputs this looks weird at a first glance (and also at a second glance) but as the configuration is in place, we can define the necessary bits and pieces for the output sharing and close this gap.

Sharing the Output between stacks

Now we sparkle some Terramate magic dust onto the stacks to connect the output of the subaccount_dev stack with the input of the cloudfoundry_dev stack.

We define a backend (not to mix up with a remote backend of Terraform) for the sharing. To do so we create a file called terramate.tm at the root level (= one level above the stack directories) with the following content:

terramate {
  config {
    experiments = [
      "outputs-sharing"
    ]
  }
}

sharing_backend "default" {
  type     = terraform
  filename = "sharing_generated.tf"
  command  = ["terraform", "output", "-json"]
}

The config block advices Terramate to enable the experimental feature of output sharing. The sharing_backend block defines which filename should be used to generate the content for the sharing namely the variables as well as the output via the filename attribute. The command attribute specifies the command Terramate should use to access the output. It's the usual suspect when it comes to outputs of Terraform.

Next, we configure Terramate what is the output that should be shared. We achieve this by adding a configuration inside of the subaccount_dev stack that we call sa_dev_config.tm with the following content:

output "cf_api_url" {
  backend   = "default"
  value     = "${jsondecode(btp_subaccount_environment_instance.cfenv_dev_base.labels)["API Endpoint"]}"
  sensitive = false
}

output "cf_org_id" {
  backend   = "default"
  value     = btp_subaccount_environment_instance.cfenv_dev_base.platform_id
  sensitive = false
}

In a nutshell we define the output variables of the stack via output blocks. The configuration tells Terramate which backend to use (the default one we created before), the value of the output as well as the sensitivity of the value.

As a counterpart we define the missing variables in the cloudfoundry_dev stack via the file cf_dev_config.tm inside of the stack directory:

input "cf_api_url" {
  backend       = "default"
  from_stack_id = "ca4662d3-b75d-4290-9842-5bb8ef924d97"
  value         = outputs.cf_api_url.value
  mock          = "https://api.cf.ap21.hana.ondemand.com"
}

input "cf_org_id" {
  backend       = "default"
  from_stack_id = "ca4662d3-b75d-4290-9842-5bb8ef924d97"
  value         = outputs.cf_org_id.value
  mock          = "917f57a1-8fee-43b3-b3a8-4bb4ce8259ab"
}

We use input blocks or the definition. Besides the backend we specify the ID of the stack where the values come from (from_stack_id) as well as the value. Beware that we are using terraform output -json, so we must reference the attribute of the JSON object via value.

UPDATE 13.12.2024: In addition, we know that the values won't be availabel during the planning phase. Hence, we add some mock data that should be used in case of an error to make the planning phase pass.

With that we are good to go and can start the code generation of Terramate via:

terramate generate

As a result, a new file will appear in the stacks named sharing_generated.tf. This corresponds to our configuration and the content of the files closes the gap of the Terraform configuration of the step before by creating the missing output

// TERRAMATE: GENERATED AUTOMATICALLY DO NOT EDIT

output "cf_api_url" {
  value = "${jsondecode(btp_subaccount_environment_instance.cfenv_dev_base.labels)["API Endpoint"]}"
}
output "cf_org_id" {
  value = btp_subaccount_environment_instance.cfenv_dev_base.platform_id
}

and input:

// TERRAMATE: GENERATED AUTOMATICALLY DO NOT EDIT

variable "cf_api_url" {
  type = any
}
variable "cf_org_id" {
  type = any
}

From a configuration perspective we are complete. Now we need to execute Terramate i.e. use Terramate to execute Terraform.

Script it!

Taking one step back, our goal is to automate the provisioning flow which means issue several Terraform commands. We do not want to do that manually firing one command at a time but we want to have a pre-defined flow for setting up and also tearing down the infrastructure.

Again, Terramate offers a solution for that namely using scripts.

As the scripting is still in experimental stage, we must activate it in the already existing configuration terramate.tm. We add the value scripts to the experiments list, so the file looks like this:

terramate {
  config {
    experiments = [
      "outputs-sharing",
      "scripts"
    ]
  }
}

sharing_backend "default" {
  type     = terraform
  filename = "sharing_generated.tf"
  command  = ["terraform", "output", "-json"]
}

Next, we define the two scripts. We call them deploy.tm and teardown.tm and store them in the stacks directory.

The deploy.tm contains the command flow from initialization, validation, planning and applying of the Terraform configuration. According to the Terramate documentation the file has the following content:

script "deploy" {
  job {
    name        = "Terraform Deployment"
    description = "Init, validate, plan, and apply Terraform changes."
    commands = [
      ["terraform", "init"],
      ["terraform", "validate"],
      ["terraform", "plan", "-out", "out.tfplan", "-lock=false", {
        enable_sharing = true
        mock_on_fail   = true
      }],
      ["terraform", "apply", "-input=false", "-auto-approve", "-lock-timeout=5m", "out.tfplan", {
        enable_sharing = true
        mock_on_fail   = true
      }],
    ]
  }
}

The structure is intuitive and consistent with other objects in Terramate. The only thing that must not be forgotten is to add the option that the output sharing is enabled when executing the commands via enable_sharing = true.
UPDATE 13.12.2024 In addition we add the option mock_on_fail = true which tells Terramate to fall back on the mock value if the real value from the dependent stack is not yet available.

The script for the teardown follows the same logic and looks like this:

script "teardown" {
  job {
    name        = "Terraform Teardown"
    description = "Destroy Terraform setup."
    commands = [
      ["terraform", "destroy", "-input=false", "-auto-approve", "-lock-timeout=5m", {
        enable_sharing = true
        mock_on_fail   = true
      }],
    ]
  }
}

Let us first check if we did things correctly and Terramate finds the scripts. We do so via

terramate script list

The output should look like this:

Looks good. Now let us check where the scripts can be applied based on their location in the directories. We do so via the following command:

terramate script tree

The output looks like this:

Also looks good. The scripts can be applied to our stacks.

The directory structure with all the implementations in place finally looks like this:

Now time for some action.

3, 2, 1 ... and Action

UPDATE 13.12.2024

Before we can start the fun part we add the required parameters in the stacks via terraform.tfvar files, namely the globalaccount in the SAP BTP-specific stack and the cf_space_managers as well as the cf_space_developers in the Cloud Foundry stack.

Having the mocking in place we can finally overcome the challenge of the dependencies and can spin up everything with one single command.

Let's bring all the bits and pieces together and spin up the infrastructure with the following commands (don't forget to set your authentication information for SAP BTP and Cloud Foundry before):

terramate script run -X deploy

After that you will have all the desired resources up and running on SAP BTP. The combined output of Terramate and Terraform in the console tells exactly what happened and helps a lot if you are running into issues.

Note I use the -X option as I am lazy and do not want to commit every time, I do some changes or fixes (and I did some over the course of setting things up). In a productive setup I would not recommend that

We can also tear things down via the second script:

terramate script run -X --reverse teardown

Mind the --reverse as otherwise things might get funny (I never forget something like that ... I heard of people doing so but not me ;-) ).

And that's it, we did it!

Where to find the code

If you want to take a closer look at the code, you find it on GitHub in this repository.

Summary

Overall, we must state there is no perfect way to solve the challenge we mentioned in the beginning. This is due to the way how Terraform works. One could work around this dependency issues for the provider configuration but it is always cumbersome and often leads to solutions that work differently when trying things out locally compared to CI/CD pipelines and are error-prone and hard to understand.

UPDATE 13.12.2024
Terramate can help us with the setup to make it concise and at least from a user perspective more streamlined. The experimental feature of output sharing as well as scripts support resolve our challenge with a at least from my perspective clean and understandable solution following the paradigm of stacks that I think is a huge strength of Terramate.

As we are using experimental features there are some points that could be improved from a documentation perspective. I could figure out everything after some time, but I think the documentation of the output sharing as well as the one for scripts might need some more love before making them GA.

What also caused some confusion is that the VS Code plugin marked the files using the experimental syntax as having errors. The corresponding error message in the pop-up really helps you with figuring out if something is missing, but I would have expected that it also recognizes the overall configuration. Not a big thing though as you can validate the setup via the Terramate commands to see if everything is where and how it should be.

As the features still evolve it would be great if the scripts would also support filtering for tags inside of the script. That would streamline the setup even more.

With that happy Terraforming and Terramating!