How to Grow Your Infrastructure
Posted on May 21, 2024 • 8 minutes • 1609 words
This topic has come up often over the years so I’m writing it down to express my current opinion of how your application and infrastructure should grow over time. I’ve roughly had the same opinion since working at Disney+ and moving to AWS and watching applications and infrastructure scale.
Every application and environment is unique so no one will be able to follow this progression exactly. But generally this is how I start anything I build, and how I recommend applications to grow–assuming they’re business applications that make money and have customers.
Start with serverless
This is a reasonable place to start for a few reasons:
- You don’t know exactly what your building
- Operational overhead will be low
- You can hide bad design with infrastructure
- You can offload functionality to services
I don’t mean you have to start with Lambda, but you should start with something that helps you figure out what your building and offload the things you don’t know early. If my company is using AWS then Lambda is fine. In GCP I like cloud run. Personal projects I usually use Cloudflare workers.
Any other get-out-of-my-way platform is a good option too. Wasm, render, fly, they’re all great. They meet the criteria of getting out of the way and helping you focus on figuring out what you’re building.
The infrastructure is going to be part of your application in a very critical way in this early stage. Embrace it. Use the k/v store even if you don’t think it’s going to scale long term.
The interesting thing about “application native clouds” (what I call providers that don’t have compute primitives) is they’re getting better at scaling. Not that long ago all of these offerings could not be taken seriously for anything beyond a static website. In 2024 you have to grow pretty large to justify moving off.
Servers/containers
At some point the serverless option is going to get either too expensive or too cumbersome. All the things that helped the platform get out of your way in the beginning suddenly get in your way when you scale out the features, integrations, the amount of people on the team, or lots of other reasons.
When serverless gets in the way some people get entrenched thinking it’s the best and will try to make it better by putting more tooling around it, adding templates, running locally, mocking services. I know those things can work for a while, but I’ve never seen them work long term as the application or team scales. At some point you need to build a monolith.
That’s where containers and servers are great! You may still be building a “microservice”, but it’s going to be bigger and architecturally simpler than the 10 functions, 3 queues, and a database you had before.
This is the point where you decide which components should be part of the infrastructure and which should be part of your application. A message queue is great, but if one function can fork a thread instead of going over the network it’s going to be easier to run locally, easier to reason about failures, easier to stablize for the second evolution of your app.
This is usually the point where you officially hire “ops” people. They’re the ones who can scale the infrastructure bits while you pile on features.
I usually recommend people build their Lambda functions with containers to start. It’s slower than native runtimes, but it’s more consistent as the size of the functions grow and it’s easier to transition out of Lambda when the time comes.
Be careful not to over complicate this step. It’s easy to think container == Kubernetes, and that’s not at all what I’m saying. Keep the infrastructure as simple as you can because you need to figure out what the application needs.
You’re going to be writing more application code in this phase because the lie of “only write business logic” is very apparent when you delete hundreds of lines of JSON glue holding your step functions and infrastructure together. The infrastructure is no longer part of your application; it’s a supporting feature but not entangled like it was in phase 1. Put your application in a container and run it in one of many container services, or stick it on a VM in an auto scaling group and don’t think about it until the next phase.
Multiple containers
This is the first time you should even think about owning any part of the orchestration. Everything before this should be as serverless or off-the-shelf as possible. Many applications don’t make it to this point, but lots of people start here.
Now is the first time you probably need to make some hard scaling choices. Your application is probably split into a few different containers, a database, and at least 2 supporting infrastructure components (e.g. cache).
This is where you can start to optimize code paths, re-write core functionality, and write that medium post about garbage collection.
Your app doesn’t fit on a single, reasonably size cloud instance so now you need to add coordination for your deployments. That is probably going to come through some form of orchestration.
Lots of enterprises start at this level. Either because they don’t allow devs to use Lambda or they have so many applications they want to build Lambda themselves.
The problem for most platform teams is it’s harder to simplify something complex than complexifying something simple. It is a lot harder to hide something like Kubernetes than it is to build cloud native cgi-bin on top of VMs.
This is the phase where a lot of companies are stuck. They see Kubernetes as the ultimate solution for application scaling with their eyes closed to what’s possible. Kubernetes is so generic they put every application in it and say it’s good enough. Some even go as far as making it the interface for everything they do (e.g. Crossplane).
BYO Cloud
Here’s where the real pain begins. Your application, or at least the code you wrote, has room to grow so you get a little more relaxed about how it should grow. It’s easier to make more services within the confines of your orchestrator so that’s obviously what happens.
The hard part is you begin to adopt external services as part of the application again. But now these external services have very specific requirements because you know what performance you expect from them. Each external service needs to have your features or you have to run it yourself, creating more responsibility.
As you adopt more external services you may start to realize you’ve lost control of your infrastructure. You spend so much time migrating components to new versions of databases, Kubernetes, and every other service you use because your cloud provider is requiring you to migrate in 60 days. You spend so much time planning and testing upgrades this is all the platform team does anymore.
They don’t have time to build features into their ~YAML templating engine~ “platform” so application teams bolt on external services. This is big enterprise problems and now you have multiple platform teams, a cloud team, and a team team just to make sure all your bases are covered.
Your company is multi-cloud and there’s an “abstraction” (a.k.a. terraform template) in the stratosphere. It’s called “ozone” because the team thought it was clever. They need to control every aspect of your self-service experience and all the benefits of the cloud have been whittled away in AppSec and SRE scoping.
At this point the application makes money. Not a little bit of “that’s cool” money, like a lot of money. Enough money to fund the next 3 quarters of AI tomfoolery without affecting your bottom line.
Reduce spend
If you’ve made it this far your application is officially Legacy with a capital “L”. Not legacy because you don’t want to write apps in Ruby any more, but Legacy because the documentation and decisions you made will out live the team who built it.
The next, and final, cycle of application evolution is to make it run bug complete with as minimal changes required and as minimal spend as possible. This is where most companies will buy 3 years of reserved instances and say that’s the cheapest they can make it.
Maybe they’ll put it in its own Kubernetes cluster just so it doesn’t interfere with the other workloads and quarterly Kubernetes upgrades can be handled separately. It will likely never have another major refactor, and anyone who’s onboarding to the stack at this point will be the caregiver, not the doctor.
Here I would recommend moving the application out of the cloud. Unless your application has a recurring 10x surge (e.g. Black Friday), most of the benefits of the cloud won’t outweigh the simplicity and cost effectiveness of static infrastructure.
Elastic scaling is cool, but it takes engineering effort and scaling rented computers up and down is still not as cheap as buying servers for 4+ years. Plus more things can, and will, go wrong when things shift around to try and optimize the last few dollars.
If you liked deleting Lambda JSON, you’ll love deleting IAM. You’ll take the 83 EC2 instances your application was running on and put it on 4 servers in a colo. Everything will be simpler because you’ve forgotten what it was like to control your environment and have real performance instead of artificial account limits.
Put the stable, Legacy application in a space where it can retire and let them. Don’t optimize it to death. Move on to the next wave of excited ideas and start the cycle over again.