Experiments in Stateless Terraform
This article about the history of statefulness in Terraform really resonated with me. Over the last few months I have been playing with setting up Terraform in such a way that the state file was no-longer required.
This post is about sharing my observations from those experiments because what I found instead was that with the current way AWS works, while it is possible to use Terraform without state, there are some limitations with AWS that made this more difficult than you would expect.
The key issue you encounter when you no-longer know the state of your infrastructure is identifying which resources are no-longer needed. You either need to enumerate them, or to discard them without being aware that they exist. The first solution I tried was to use the AWS Account as an ephemeral container for everything that is built and destroyed. This guarantees to find everything and is reasonably quick to achieve.
I do recognise that what I’m attempting to do is a hack, Terraform and AWS are not designed to be used this way. Terraform has not made their decision to use state in isolation, they will have encountered the problems I further on in this article and the solution they went for was adding complexity through state. I don’t see that as unreasonable, they made those decisions based upon the limitations of AWS at the time, and AWS made their own decisions based on assumptions about how their infrastructure was intended to be used.
The concept is that everything you need is first built in to one AWS Account and when you want to release a new instance, re-create everything in a second AWS Account and fail over to it. The first AWS Account can then be removed. A separate AWS Account with a load balancer can then perform the switching when the new infrastructure is up and running. This is a form of Green/Blue Deployment.
There are a some limitations with this, and this is where I think it gets interesting.
My first thought to clearing out the old resources is to delete the AWS Account. However, the deletion of AWS Accounts from an AWS Organization requires the sub-account to have credit card details associated to the Account. That makes it a manual step. A manual step means no automated Green/Blue deployment.
There is an alternative. I call it the “black hole”. This is an OU in your AWS Organization that has a policy on it that prevents all roles from being assumed in that Account. That way your resources cannot run and your cost will reach zero.
This then runs in to another issue. There are soft limits on the number of sub-Accounts that an AWS Organization can have. By default 10. When you’re doing continuous deployments, you do the math on the number of Accounts you will need per day. It is a soft limit, so you can ask Amazon nicely for it to be increased, but I bet in a short amount of time you will have an unhappy Amazon asking you what you are up to and can you please stop.
The only other way of doing this is to follow the enumeration path. Fortunately tooling such as cloud-nuke exist to empty an AWS Account of resources. You can even be selective about allow-listing certain resources, but it is a very slow process and it may not cover everything.
However, the interesting part here was finding the limits of what was possible in AWS. Not something I encounter often. I don’t have a lot of experience in other PaaS services so I would love to hear if doing this is possible in Azure or GCP.
I do wonder if its something Amazon will fix, but I also think it is a fundamental limitation in how AWS assumed Accounts would be used. Otherwise they would have no doubt made it easier.
A different, but unattainable solution I’ve started to see with services like Fastly and Doppler is that they have configuration versioning built right in to their web UI.
I am hopeful that this is a tend towards a ClickOps model which has the ability to make Infrastructure as Code, and therefore managing state, redundant. The first PaaS services to do infrastructure versioning this way is going to get a lot of attention to me. However, I don’t see this happening in AWS for a long time.