In the previous post, Robert introduced Wild Workouts, our example serverless application. Every week or two, we will release new articles related to this project, focusing on creating business-oriented applications in Go. In this post, I continue where Robert left off and describe the infrastructure setup.

We picked Google Cloud Platform (GCP) as the provider of all infrastructure parts of the project. We use Cloud Run for running Go services, Firestore as the database, Cloud Build as CI/CD, and Firebase for web hosting and authentication. The project is based on Terraform.

Infrastructure as Code

If you’re not familiar with Terraform at all, it’s a tool for keeping infrastructure configuration as text files. This technique is also known as “Infrastructure as Code”. It’s a broad topic, so I’ll mention just a few benefits:

  • Storing infrastructure configuration in a repository. This gives you all benefits that come with version control, like the history of commits, one source of truth, and code reviews.
  • Consistency over time. Ideally, there are no manual changes to infrastructure other than those described in the configuration.
  • Multiple environments. When you stop treating your servers as pets, it’s much easier to create identical environments, sometimes even on demand.

Terraform 101

You can skip this part if you already know Terraform basics.

In Terraform, you define a set of resources in .tf files using HCL syntax. A resource can be a database, network configuration, compute instance, or even a set of permissions. Available resource types depend on the used provider (e.g., AWS, GCP, Digital Ocean).

All resources are defined in a declarative way. You don’t write pseudocode that creates servers. Instead, you define the desired outcome. It’s up to Terraform to come up with a way to configure what you specified. For example, it’s smart enough to create a resource only if it doesn’t exist.

Resources can refer to each other using the full name.

Besides resources, a project can also specify input variables to be filled by user and output values, that can be printed on the screen or stored in a file.

There are also data sources that don’t create anything but read existing remote resources.

You can find all the available resources in the specific provider documentation. Here’s an example for Google Cloud Platform.

Two basic commands that you need to know are terraform plan and terraform apply.

apply applies all resources defined in the current directory in all files with .tf extension. The plan command is a “dry-run” mode, printing out all changes that would be applied by apply.

After applying changes, you will find a terraform.tfstate file in the same directory. This file holds a local “state” of your infrastructure.

Google Cloud Project

Our Terraform configuration creates a new GCP project. It’s completely separated from your other projects and is easy to clean up.

Because some resources go beyond the free trier, you need to have a Billing Account. It can be an account with a linked credit card, but the $300 credit for new accounts is also fine.

The basic part of the project definition looks like this:

provider "google" {
  project = var.project
  region  = var.region

data "google_billing_account" "account" {
  display_name = var.billing_account

resource "google_project" "project" {
  name            = "Wild Workouts"
  project_id      = var.project
  billing_account =

resource "google_project_iam_member" "owner" {
  role    = "roles/owner"
  member  = "user:${var.user}"

  depends_on = [google_project.project]

Let’s consider what each block does:

  1. Enables the GCP provider. Sets project name and chosen region from variables. These two fields are “inherited” by all resources unless overridden.
  2. Finds a billing account with the display name provided by the variable.
  3. Creates a new google project linked with the billing account. Note the reference to the account ID.
  4. Adds your user as the owner of the project. Note the string interpolation.

Enabling APIs

On a fresh GCP project, you can’t use most services right away. You have to first enable the API for each of them. You can enable an API by clicking a button in the GCP Console, or do the same in Terraform:

resource "google_project_service" "compute" {
  service    = "[]("
  depends_on = [google_project.project]

We enable the following APIs:

  • Cloud Source Repositories
  • Cloud Build
  • Container Registry
  • Compute
  • Cloud Run
  • Firebase
  • Firestore

A note on dependencies

If you’re wondering about the depends_on line, it’s there to set an explicit dependency between the service and the project. Terraform detects dependencies between resources if they refer to each other. In the first snippet, you can see this with billing account that’s referenced by the project:

data "google_billing_account" "account" {
  display_name = var.billing_account

resource "google_project" "project" {
  # ...
  billing_account =

In the google_project_service resource, we don’t use project anywhere, because it’s already set in the provider block. Instead, we use depends_on to specify explicit dependency. This line tells Terraform to wait with creating the resource until the project is correctly created.

Cloud Run

Cloud Run services

In Cloud Run, a service is a set of Docker containers with a common endpoint, exposing a single port (HTTP or gRPC). Each service is automatically scaled, depending on the incoming traffic. You can choose the maximum number of containers and how many requests each container can handle.

It’s also possible to connect services with Google Cloud Pub/Sub. Our project doesn’t use it yet, but we will introduce it in the future versions.

You are charged only for the computing resources you use, so when a request is being processed or when the container starts.

Wild Workouts consists of 3 services: trainer, trainings, and users. We decided to serve public API with HTTP and internal API with gRPC.


Because you can’t have two separate ports exposed from a single service, we have to expose two containers per service (except trainings, which doesn’t have internal API at the moment).

We deploy 5 services in total, each with a similar configuration. Following the DRY principle, the common Terraform configuration is encapsulated in the service module.

The module holds a definition of a single generic Cloud Run service and is used multiple times in It exposes several variables, e.g., name and type of the server (HTTP or gRPC).

A service exposing gRPC is the simpler one:

module cloud_run_trainer_grpc {
  source = "./service"

  project    = var.project
  location   = var.region
  dependency = null_resource.init_docker_images

  name     = "trainer"
  protocol = "grpc"

The protocol is passed to the SERVER_TO_RUN environment variable, which decides what server is run by the service.

Compare this with an HTTP service definition. We use the same module, and there are additional environment variables defined. We need to add them because public services contact internal services via the gRPC API and need to know their addresses.

module cloud_run_trainings_http {
  source = "./service"

  project    = var.project
  location   = var.region
  dependency = null_resource.init_docker_images

  name     = "trainings"
  protocol = "http"
  auth     = false

  envs = [
      name = "TRAINER_GRPC_ADDR"
      value = module.cloud_run_trainer_grpc.endpoint
      name  = "USERS_GRPC_ADDR"
      value = module.cloud_run_users_grpc.endpoint

The reference: module.cloud_run_trainer_grpc.endpoint points to endpoint output defined in the service module:

output endpoint {
  value = "${trimprefix(google_cloud_run_service.service.status[0].url, "https://")}:443"

Using environment variables is an easy way to make services know each other. It would probably be best to implement some kind of service discovery with more complex connections and more services. Perhaps we will cover this in another post in the future.

If you’re curious about the dependency variable see In-depth down below for details.

Cloud Run Permissions

Cloud Run services have authentication enabled by default. Setting the auth = false variable in the service module adds additional IAM policy for the service, making it accessible for the public. We do this just for the HTTP APIs.

data "google_iam_policy" "noauth" {
  binding {
    role = "roles/run.invoker"
    members = [

resource "google_cloud_run_service_iam_policy" "noauth_policy" {
  count = var.auth ? 0 : 1

  location = google_cloud_run_service.service.location
  service  =

  policy_data = data.google_iam_policy.noauth.policy_data

Note the following line:

count = var.auth ? 0 : 1

This line is Terraform’s way of making an if statement. count defines how many copies of a resource Terraform should create. It skips resources with count equal 0.


We picked Firestore as the database for Wild Workouts. See our first post for reasons behind this.

Firestore works in two modes - Native or Datastore. You have to decide early on, as the choice is permanent after your first write to the database.

You can pick the mode in GCP Console GUI, but we wanted to make the setup fully automated. In Terraform, there’s google_project_service resource available, but it enables the Datastore mode. Sadly, we can’t use it, since we’d like to use the Native mode.

A workaround is to use the gcloud command to enable it (note the alpha version).

gcloud alpha firestore databases create

To run this command, we use the null resource. It’s a special kind of resource that lets you run custom provisioners locally or on remote servers. A provisioner is a command or other software making changes in the system.

We use local-exec provisioner, which is simply executing a bash command on the local system. In our case, it’s one of the targets defined in Makefile.

resource "null_resource" "enable_firestore" {
  provisioner "local-exec" {
    command = "make firestore"

  depends_on = [google_firebase_project_location.default]

Firestore requires creating all composite indexes upfront. It’s also available as a Terraform resource.

resource "google_firestore_index" "trainings_user_time" {
  collection = "trainings"

  fields {
    field_path = "UserUuid"
    order      = "ASCENDING"

  fields {
    field_path = "Time"
    order      = "ASCENDING"

  fields {
    field_path = "__name__"
    order      = "ASCENDING"

  depends_on = [null_resource.enable_firestore]

Note the explicit depends_on that points to the null_resource creating database.


Firebase provides us with frontend application hosting and authentication.

Currently, Terraform supports only part of Firebase API, and some of it is still in beta. That’s why we need to enable the google-beta provider:

provider "google-beta" {
  project     = var.project
  region      = var.region
  credentials = base64decode(google_service_account_key.firebase_key.private_key)

Then we define the project, project’s location and the web application.

resource "google_firebase_project" "default" {
  provider = google-beta

  depends_on = [

resource "google_firebase_project_location" "default" {
  provider = google-beta

  location_id = var.firebase_location

  depends_on = [

resource "google_firebase_web_app" "wild_workouts" {
  provider     = google-beta
  display_name = "Wild Workouts"

  depends_on = [google_firebase_project.default]

Authentication management still misses a Terraform API, so you have to enable it manually in the Firebase Console. Firebase authentication is the only thing we found no way to automate.

Cloud Run services

Firebase Routing

Firebase also handles public routing to services. Thanks to this the frontend application can call API with /api/trainer instead of These routes are defined in web/firebase.json.

"rewrites": [
      "source": "/api/trainer{,/**}",
      "run": {
        "serviceId": "trainer-http",
        "region": "europe-west1"
    // ...

Cloud Build

Cloud Build is our Continuous Delivery pipeline. It has to be enabled for a repository, so we define a trigger in Terraform and the repository itself.

resource "google_sourcerepo_repository" "wild_workouts" {
  name = var.repository_name

  depends_on = [

resource "google_cloudbuild_trigger" "trigger" {
  trigger_template {
    branch_name = "master"
    repo_name   =

  filename = "cloudbuild.yaml"

The build configuration has to be defined in the cloudbuild.yaml file committed to the repository. We defined a couple of steps in the pipeline:

  • Linting (go vet) - this step could be extended in future with running tests and all kinds of static checks and linters
  • Building docker images
  • Deploying docker images to Cloud Run
  • Deploying web app to Firebase hosting

We keep several services in one repository, and their building and deployment are almost the same. To reduce repetition, there are a few helper bash scripts in the scripts directory.

As usual in CI/CD tools, the Build configuration is defined in YAML. Let’s see the whole configuration for just one service, to reduce some noise.

- id: trainer-lint
  name: golang
  entrypoint: ./scripts/
  args: [trainer]

- id: trainer-docker
  entrypoint: ./scripts/
  args: ["trainer", "$PROJECT_ID"]
  waitFor: [trainer-lint]

- id: trainer-http-deploy
  entrypoint: ./scripts/
  args: [trainer, http, "$PROJECT_ID"]
  waitFor: [trainer-docker]

- id: trainer-grpc-deploy
  entrypoint: ./scripts/
  args: [trainer, grpc, "$PROJECT_ID"]
  waitFor: [trainer-docker]

  - 'GO111MODULE=on'
  machineType: 'N1_HIGHCPU_8'

- '$PROJECT_ID/trainer'

A single step definition is pretty short:

  • id is a unique identifier of a step. It can be used in waitFor array to specify dependencies between steps (steps are running in parallel by default).
  • name is the name of a docker image that will be run for this step.
  • entrypoint works like in docker images, so it’s a command that is executed on the container’s start. We use bash scripts to make the YAML definition short, but this can be any bash command.
  • args will be passed to entrypoint as arguments.

In options, we override the machine type to make our builds run faster. There’s also an environment variable to force the use of go modules.

images list defines docker images that should be pushed to Container Registry. In the example above, the docker image is built in the trainer-docker step.


Deploys to Cloud Run are done with the gcloud run deploy command. Cloud Build builds the docker image in the previous step, so we can use the latest image on the registry.

- id: trainer-http-deploy
  entrypoint: ./scripts/
  args: [trainer, http, "$PROJECT_ID"]
  waitFor: [trainer-docker]
gcloud run deploy "$service-$server_to_run" \
    --image "$project_id/$service" \
    --region europe-west1 \
    --platform managed

Frontend is deployed with the firebase command. There’s no need to use a helper script there, as there’s just one frontend application.

- name:$PROJECT_ID/firebase
  args: ['deploy', '--project=$PROJECT_ID']
  dir: web
  waitFor: [web-build]

This step uses$PROJECT_ID/firebase docker image. It doesn’t exist by default, so we use another null_resource to build it from cloud-builders-community:

resource "null_resource" "firebase_builder" {
  provisioner "local-exec" {
    command = "make firebase_builder"

  depends_on = [google_project_service.container_registry]

Cloud Build Permissions

All required permissions are defined in Cloud Build needs permissions for Cloud Run (to deploy backend services) and Firebase (to deploy frontend application).

First of all, we define account names as local variables to make the file more readable. If you’re wondering where these names come from, you can find them in the Cloud Build documentation.

locals {
  cloud_build_member = "serviceAccount:${google_project.project.number}"
  compute_account    = "projects/${var.project}/serviceAccounts/${google_project.project.number}"

We then define all permissions, as described in the documentation. For example, here’s Cloud Build member with Cloud Run Admin role:

resource "google_project_iam_member" "cloud_run_admin" {
  role   = "roles/run.admin"
  member = local.cloud_build_member

  depends_on = [google_project_service.cloud_build]


We have a couple of Dockerfiles defined in the project in the docker directory.

  • app - Go service image for local development. It uses reflex for hot code recompilation. See our post about local Go environment to learn how it works.
  • app-prod - production image for Go services. It builds the Go binary for a given service and runs it.
  • web - frontend application image for local development.


Usually, you would execute a Terraform project with terraform apply. This command applies all resources in the current directory.

Our example is a bit more complex because it sets up the whole GCP project along with dependencies. Makefile orchestrates the setup.


  • git, gcloud, and terraform installed on your system.
  • Docker
  • A GCP account with a Billing Account.


There are two credentials that you need for the setup.

Start with logging in with gcloud:

gcloud auth login

Then you need to obtain Application Default Credentials:

gcloud auth application-default login

This stores your credentials in a well-known place, so Terraform can use it.

Picking a Region

During the setup, you need to pick Cloud Run region and Firebase location. These two are not the same thing (see the lists on linked documentation).

We picked europe-west1 as the default region for Cloud Run. It’s hardcoded on the repository in two files:

If you’re fine with the defaults, you don’t need to change anything. Just pick Firebase location that’s close, i.e., europe-west.

If you’d like to use another region, you need to update the files mentioned above. It’s important to commit your changes on master before you run the setup. It’s not enough to change them locally!

Run it!

Clone the project repository and enter the terraform directory from the command line. A single make command should be enough to start. See the included README for more details.

At the end of the setup, a new git remote will be added to your local repository, called google. The master branch will be pushed to this remote, triggering your first Cloud Build.

If you’d like to make any changes to the project, you need to push to the correct origin, i.e., git push google. You can also update the origin remote with git remote set-url.


Env variables

All possible variables are defined in the file. If a variable has no default value and you don’t provide it, Terraform won’t let you apply anything.

There are several ways to supply variables for terraform apply. You can pass them individually with -var flag or all at once with -var-file flag. Terraform also looks for them in TF_VAR_name environment variables.

We’ve picked the last option for this project. We need to use the environment variables anyway in Makefile and bash scripts. This way, there’s just one source of truth.

Make runs at the beginning of the setup. It’s just a helper script to ask the user for all the required variables. These are then saved to .env file. You can edit this file manually as well.

Note Terraform does not automatically read it. It’s sourced just before running apply:

load_envs:=source ./.env
# ...

	${load_envs} && terraform apply

Building docker images

Because we’re setting up the project from scratch, we run into a dilemma: we need docker images in the project’s Container Registry to create Cloud Run instances, but we can’t create the images without a project.

You could set the image name in Terraform definition to an example image and then let Cloud Build overwrite it. But then each terraform apply would try to change them back to the original value.

Our solution is to build the images based on the “hello” image first and then deploy Cloud Run instances. Then Cloud Build builds and deploys proper images, but the name stays the same.

resource "null_resource" "init_docker_images" {
  provisioner "local-exec" {
    command = "make docker_images"

  depends_on = [google_project_service.container_registry]

Note the depends_on that points at Container Registry API. It ensures Terraform won’t start building images until there’s a registry where it can push them.


You can delete the entire project with make destroy. If you take a look in the Makefile, there’s something unusual before terraform destroy:

terraform state rm "google_project_iam_member.owner"
terraform state rm "google_project_service.container_registry"
terraform state rm "google_project_service.cloud_run"

These commands are a workaround for Terraform’s lack of a “skip destroy” feature. They remove some resources from Terraform’s local state file. Terraform won’t destroy these resources (as far as it’s concerned, they don’t exist at this point).

We don’t want to destroy the owner IAM policy, because if something goes wrong during the destroy, you will lock yourself out of the project and won’t be able to access it.

The other two lines are related with enabled APIs — there’s a possible race condition where some resources would still use these APIs during destroy. By removing them from the state, we avoid this issue.

A note on magic

Our goal was to make it as easy as possible to set up the whole project. Because of that, we had to cut some corners where Terraform is missing features, or an API is not yet available. There’s also some Makefile and bash magic involved that’s not always easy to understand.

I want to make this clear because you will likely encounter similar dilemmas in your projects. You will need to choose between fully-automated solutions glued together with alpha APIs or using plain Terraform with some manual steps documented in the project.

Both approaches have their place. For example, this project is straightforward to set up multiple times with the same configuration. So if you’d like to create the exact copy of the production environment, you can have a completely separate project working within minutes.

If you keep just one or two environments, you won’t need to recreate them every week. It’s then probably fine to stick to the available APIs and document the rest.

Is serverless the way to go?

This project is our first approach to “serverless” deployment, and at first, we had some concerns. Is this all just hype, or is serverless the future?

Wild Workouts is a fairly small project that might not show all there is to Cloud Run. Overall we found it quite simple to set up, and it nicely hides all complexities. It’s also more natural to work with than Cloud Functions.

After the initial setup, there shouldn’t be much infrastructure maintenance needed. You don’t have to worry about keeping up a Docker registry or a Kubernetes cluster and can instead focus on creating the application.

On the other hand, it’s also quite limited in features, as there are only a few protocols supported. It looks like a great fit for services with a REST API. The pricing model also seems reasonable, as you pay only for the used resources.

What about vendor lock-in?

The entire Terraform configuration is now tied to GCP, and there is no way around it. If we’d like to migrate the project to AWS, a new configuration will be needed.

However, there are some universal concepts. Services running in docker images can be deployed on any platform, whether it’s a Kubernetes cluster or docker-compose. Most platforms also offer some kind of registry for images.

Is writing a new Terraform configuration all that’s required to migrate out of Google? Not really, as the application itself is coupled to Firebase and Authentication offered by GCP. We will show a better approach to this problem in future posts in this series.

What’s next?

This post just scratches the surface of Terraform. There are many advanced topics you should take a look at if you consider using it. For example, the remote state is necessary when working in a team.

I hope this was a solid introduction. If you have any questions, please let us know in the comments below. We might even follow up with full posts to some of them, as the topic is broad enough.

Next week, we’re going to cover gRPC communication. 👋