Self-referential deployment pointers

Here is cool trick you can do in Terraform in order to remember last version of the application you deployed.

Common way to do it is to declare a variable

First of all, you need an variable to read version from

variable "release_version" {
  type    = string
  default = null
}

And then have a tfvars file in the source code repository, which will point to the current version of the application:

release_version = "0cf879d3809e428787f35ace54b45c6f"

Having that in place is convenient for many reasons: - it is immediately obvious which version is currently deployed and who performed the deploy (git log holds the history) - repeated invocations of terraform apply are idempotent, no need to worry about “forgetting” to specify -var cli argument

But there is a different way, which might be more convenient in case you want to deploy very often, in automated way (from CI, without human interaction).

Declare an output, which captures a version of the application after deploy:

module "application" {
  source = "../modules/application"
  image_tag = coalesce(var.release_version, data.terraform_remote_state.self.outputs.deployed_release_version)
}

output "release_version" {
  value = module.application.tagged_image.source_tag
}

And a terraform_remote_state data block, which points to the terraform state of the project itself (hence the self name).

data "terraform_remote_state" "self" {
  backend = "s3"
  config = {
    bucket = "terraform-state"
    key    = "key.tfstate"
  }
}

The magic sauce is this line, which tries to read from input variable, and, if absent, from the output of self state.

coalesce(var.release_version, data.terraform_remote_state.self.outputs.deployed_release_version)

So we have idempotency for free (it is safe to invoke terraform apply without parameters) and ability to release a new version by providing a -var=release_version=new_sha cli argument to terraform apply