How to organize terraform modules for multiple environments?

Question

Every Terraform guide on the web provides a partial solution that is almost always not the real picture.
I get that, not everyone has the same infrastructure needs, but what worries me that the common scenario with:

multiple environments (dev, stage)
remote backend (s3)
some basic resources (bucket or ec2 instance)

isn't presented anywhere on a real example project.
I'm looking for just that, and in the meantime, I have researched and concluded that apart from those needs I also want:

to utilize modules
to NOT use workspaces, but rather a distinct directory-per-environment approach
to NOT use terragrunt wrapper

My current structure, which does not utilize modules - only root module:

infra/ ------------------------------ 'terraform init', 'terraform apply' inside here*  
     main.tf ------------------------ Sets up aws provider, backend, backend bucket, dynamodb table   
     terraform.tfvars   
     variables.tf ----------------- Holds few variables such as aws_region, project_name...

My desired structure folder tree (for a simple dev & staging simulation of a single bucket resource) is I think something like this:

infra/  
     dev/  
        s3/  
            modules.tf ------ References s3 module from local/remote folder with dev inputs   
     stage/  
        s3/  
            modules.tf ------ References s3 module from local/remote folder with stage inputs

But what about the files from my previous root module? I still want to have a remote backend in the same way as before, just now I want to have two state files (dev.tfstate and stage.tfstate) in the same backend bucket? How would the backend.tf files look like in each subdirectory and where would they be? In s3/ folder or dev/ folder?

It's kind of confusing since I'm transitioning from root module 'terraform init' approach, to specific subdirectory 'terraform init', and it's not clear to me whether I should still have a root-module or another folder for example called global/ which I should consider my prerequisite which I should init at the beginning of the project and is basically leave alone from that point on since it created the buckets which dev/ and staging/ can reference?

One more question is: what if I have s3/ ec2/ ecr/ subdirectories inside each environment, where do I execute 'terraform plan' command? Does it traverse all subdirectories?

When I have the answers and a clear picture of this above, it would be great to improve it by DRYing it up, but for now, I value a more practical solution through example rather than just a theoretic DRY explanation. Thanks!

You do not want different modules.tf for different stages, that is exactly NOT how your infrastructure is supposed to look like. Because now your stage no longer is related to your dev and what you develop on dev will not be the same thing tested on stage will not be the same thing released on prod. You only have one directory where you put all the infrastrucutre, all the modules, all the logic. This is parametrized by some variables, e.g. the ec2 instance size or log retention time, etc. Those variables will have different values depending on the env, that is the only thing that changes. — luk2302
And then you WILL need a wrapper, either terragrunt or some custom basic shell script. Technically you are not required to have a wrapper but the alternative is having to remember a lot of parameters for the terraform commands. — luk2302
Your findings seem to be the exact opposite of what I would recommend. Terraform workspaces exist to solve the exact problem you are trying to solve. Why do you think workspaces won't work for you? — Mark B
How are workspaces "anti-documenting"? You should really play around with these features some I think, to get an idea of how they work. Workspaces and modules are two totally different things, that work perfectly fine together. Modules should be used to encapsulate a set of resources, like a VPC module, an EC2 module, etc. I highly recommend browsing (or even using) some of the open source modules in the Terraform registry to see how they work registry.terraform.io/browse/modules — Mark B
@MarkB I have no trouble understanding how a module works. Heck I even created one s3 module on separate repo and I am aware of how to call it. Most of my questions come from separating stuff into subdirectories and then not knowing how to manage state from that point on. Simply I have trouble understanding how I should configure and manage state in this subdirectories fashion. MANY examples on web talk about how you “shouldn’t put everything in one place” but FEW talk about how your terraform commands should adapt to acompany that. — deltakroneker

Daniel Hornik Daniel Hornik · Accepted Answer · 2021-02-03T10:48:44

I work with terraform 5 years. I did a lot of mistakes with in my career with modules and environments. Below text is just share of my knowledge and experience. They may be bad.

Real example project may is hard to find because terraform is not used to create opensource projects. It's often unsafe to share terraform files because you are showing all vulnerabilities from your intrastructure

Module purpose and size

You should create module that has single purpose, but your module should be generic.

Example module

You can create bastion host module, but better idea is to create a module for generic server. This module may have some logic dedicated to your business problem like, CW Log group, some generic security group rules, etc.

Application module

Sometimes it is worth to create more specific module.

Let's say you have application, that requires Lambda, ECS service, CloudWatch alarms, RDS, EBS etc. All of that elements are strongly connected.

You have 2 options:

Create separated modules for each above items - But then your application uses 5 modules.
Create one big module and then you can deploy your app with single module
Mix above solutions - I prefer that

Everything depends on details and some circumstances.

But I will show you how I use terraform in my productions in different companies.

Separated definitions for separated resurces

This is project, where you have environment as directories. For each application, networking, data resoruces you have separated state. I keep mutable data in separated directory(like RDS, EBS, EFS, S3, etc) so all apps, networking, etc can be destroyed and recreated, because they are stateless. No one can destroy statefull items because data can be lost. This is what i was doing for last few years.

project/
├─ packer/
├─ ansible/
├─ terraform/
│  ├─ environments/
│  │  ├─ production/
│  │  │  ├─ apps/
│  │  │  │  ├─ blog/
│  │  │  │  ├─ ecommerce/
│  │  │  ├─ data/
│  │  │  │  ├─ efs-ecommerce/
│  │  │  │  ├─ rds-ecommerce/
│  │  │  │  ├─ s3-blog/
│  │  │  ├─ general/
│  │  │  │  ├─ main.tf
│  │  │  ├─ network/
│  │  │  │  ├─ main.tf
│  │  │  │  ├─ terraform.tfvars
│  │  │  │  ├─ variables.tf
│  │  ├─ staging/
│  │  │  ├─ apps/
│  │  │  │  ├─ ecommerce/
│  │  │  │  ├─ blog/
│  │  │  ├─ data/
│  │  │  │  ├─ efs-ecommerce/
│  │  │  │  ├─ rds-ecommerce/
│  │  │  │  ├─ s3-blog/
│  │  │  ├─ network/
│  │  ├─ test/
│  │  │  ├─ apps/
│  │  │  │  ├─ blog/
│  │  │  ├─ data/
│  │  │  │  ├─ s3-blog/
│  │  │  ├─ network/
│  ├─ modules/
│  │  ├─ apps/
│  │  │  ├─ blog/
│  │  │  ├─ ecommerce/
│  │  ├─ common/
│  │  │  ├─ acm/
│  │  │  ├─ user/
│  │  ├─ computing/
│  │  │  ├─ server/
│  │  ├─ data/
│  │  │  ├─ efs/
│  │  │  ├─ rds/
│  │  │  ├─ s3/
│  │  ├─ networking/
│  │  │  ├─ alb/
│  │  │  ├─ front-proxy/
│  │  │  ├─ vpc/
│  │  │  ├─ vpc-pairing/
├─ tools/

To apply single application, You need to do:

cd ./project/terraform/environments/<ENVIRONMENT>/apps/blog;

terraform apply;

You can see there is a lot of directories in all environments. As I can see there are pros and cons of that tools.

Cons:

It is hard to check if all modules are in sync
Complicated CI
Complicated directory structure especially for new people in the team, but it is logic
There may be a lot of dependencies, but this is not a problem when you think about it from the beginning.
You need to take care, to keep exactly the same environments.
There is a lot of initialization required and refactors are hard to do.

Pros:

Quick apply after small changes
Separated applications and resources. It is easy to modify small module or small deployment for it without knowledge about overall system
It is easier to clean up when you remove something
It's easy to tell what module need to be fixed. I use some tools I wrote to analyze status of particular parts of infrastructure and I can send email to particular developer, that his infrastructure needs resync for some reasons.
You can have different environments easier than in the monolith. You can destroy single app if you do not need it in environemnt

Monolith infrastructure

Last time I started working with new company. They keep infrastructure definition in few huge repositories(or folders), and when you do terraform apply, you create all applications at the same time.

project/
├─ modules/
│  ├─ acm/
│  ├─ app-blog/
│  ├─ app-ecommerce/
│  ├─ server/
│  ├─ vpc/
├─ vars/
│  ├─ user/
│  ├─ prod.tfvars
│  ├─ staging.tfvars
│  ├─ test.tfvars
├─ applications.tf
├─ providers.tf
├─ proxy.tf
├─ s3.tf
├─ users.tf
├─ variables.tf
├─ vpc.tf

Here you prepare different input values for each environment.

So for example you want to apply changes to prod:

terraform apply -var-file=vars/prod.tfvars -lock-timeout=300s

Apply staging:

terraform apply -var-file=vars/staging.tfvars -lock-timeout=300s

Cons:

You have no dependency, but sometimes you need to prepare some environment element like domains, elastic IP, etc manually, or you need to have them created before terraform plan/apply. Then you have problem
Its hard to do cleanup as you have hundreds resources and modules at the same time
Extremely long terraform execution. Here it takes around 45 minutes to plan/apply single environment
It's hard to understand entire environment.
Usually you need to have 2/3 repositories if you keep that structure to separate networking,apps,dns etc...
You need to do much more work to deal with different environments. You need to use count etc...

Pros:

It's easy to check if your infrastructure is up to date
There is no complicated directory structure...
All your environments are exactly the same.
Refactoring may be easier, because you have all resources in very few places.
Small number of initialization is required.

Summary

As you can see this is more architectural problem, the only way to learn it, is to get more experience or read some posts from another people...

I am still trying to figure out the most optimal way and I would probably experiment with first way.

Do not take my advantages as sure thing. This post is just my experience, maybe not the best.

References

I will post some references that helped me a lot: