Much of my system design philosophy was forged during my time at Facebook. We did a lot of things differently that I think was responsible for our technical success. I want to share one of them with you. It’s the most common pitfall I've see prevalent in our industry, and it relates how we solve problems at the system level. I’ll illustrate it with a story:
You kick off your startup. One service, one repo, one database. Version zero ships quickly. You iterate and iterate until you sense product market fit.
All of a sudden, you’ve got it and you’re on a rocket ship. You grow your team, you’re onboarding customers, and you’ve got some great problems on your hands.
One of those problems: your systems break.
Your build infrastructure slows down. Deployment becomes an all-day, panic-ridden affair. Commits start to break unrelated components more and more frequently. The development environment becomes slower and slower. You get incidents left and right. Oncall becomes more demanding than a newborn at night.
Slow down productivity and you’ll get grumpy engineers, but take away their sleep and you’ll get rioting engineers. Now you have an emergency on your hands.
You look around, and the technical debt is overwhelming. You think: it’s time to grow up. You research and land on microservices as the answer. It seems to solve all your biggest pain points:
Speed up deployments
Teams can manage their own deployments. Because deploying smaller systems is easier and safer than larger ones, your teams can now deploy faster. Heck, they could even roll back now.
Separate concerns
You’ll force stronger code boundaries. No more commits breaking unrelated components
Empower teams
You’ll be able to use the right tools for the job. No more need to have the same language or stack. Teams can use whatever will make the most productive at solving their particular problem.
Ensure quality
You’ll be able to clarify ownership. No more spaghetti code because service owners will ultimately be responsible
This is the dream. You kick off that initiative. But…you see some serious costs:
Hardened boundaries are harder to change
Those strict code boundaries come with their own costs. If you split incorrectly, you’re prone to dupe your data and your logic. But, even if you split correctly, what happens when your business needs change? What if you get new regulation that requires system-wide changes? All of a sudden, the way you split those services won’t make sense anymore, and evolving those boundaries become a magnitude more difficult.
Hardened boundaries make it more difficult to ship unanticipated changesets
What if you need to make changes that your org and service structure didn’t anticipate? These changes won’t cleanly fit into one service, or even groups of services. All of a sudden a change that could have been done by one engineer before requires teams to align to deliver.
Hardened boundaries reduce the potential impact of your engineers
What if one of your engineers comes up with a performance improvement, or a new idea for the business? Their impact is now constrained by the services they have control over. If the languages are different they won’t be able to actualize that performance improvement for the whole company. Even if we force all services to use the same language, delivering that to all services becomes a difficult affair.
More Incidental complexity
Sharing code, managing deploys, logging infrastructure, service orchestration, rpc are all made more difficult by a change to multiple services. None of the added difficulty helps you move faster or ship with higher quality.
Why so many unanticipated issues? Because microservices as a strategy is a blanket solution.
Blanket solutions don’t specifically address any one issue, but try to address a multitude of issues. This is often done with philosophy (new principles we will follow to build things) then with technology (how exactly we will solve problems). Whenever a solution addresses a multitude of issues with philosophy, it’s likely to come with a multitude of unanticipated problems.
The alternative path is evolution. As a rule of thumb, the changes you introduce should concretely solve the problems you care about. This doesn’t mean that all solutions need to be iterative, but it does mean that all solutions need to be under strong selection pressure. Every solution should have immediate wins in sight.
For example:
Deployment, CI, logging issues?
What if you built a centralized team that owned that complexity, and built infrastructure so product engineers didn’t have to worry about it? For example, most product engineers did not have to worry about deploys, observability, or logging at Facebook. Infra was already in place that they could leverage, managed by an underlying team.
Code boundary issues?
What if you evolved your system to use immutable structures and smaller interfaces between boundaries? What if you pushed some of these problems down the stack? For example, there may be a lot of complexity introduced with privacy. You can centralize that concern, into building infrastructure with a small interface, that product engineers can use. Facebook did this with viewer context
Scaling issues?
This may be a true concern, but instead of applying a general philosophy, could we focus on solving this problem directly? Do all systems need to scale independently, or does it only matter that just a few things scale independently? exactly are the problems? Could you address with only changing the hottest, most intensive paths?
Complexity issues?
Where is the complexity exactly? Could we evolve modules so you could localize the reasoning behind them? Could we abstract the difficult portions? Sophie Alpert gives a great example of this here
This kind of thinking leads to a philosophy of system design based on simplicity:
You address problems concretely with a view towards evolving your system. At each step of the way, you constantly optimize for engineering velocity. Engineering velocity is a great metric to use, because it implies correctness and quality alongside with speed. It’s impossible to ship if you don’t have confidence in your system. You push and centralize complexity down the stack. You evolve your system so product engineers can think locally within the module they’re working on. You make changes that empower any engineer to drive impact throughout the stack.
Doing this won’t be a magic pill — it will look like your system is constantly broken and in need of improvement — but that is its secret weapon. You’re constantly evolving it.
Thanks to Daniel Woelfel, Jacky Wang, Kam Leung, Joe Averbukh, Phil Nachum, Alexandre Lebrun, for reviewing drafts of this essay