Most of you, if you were students, wrote essays or something like that as undergraduates, right? Fairly confident to say that nobody’s actually kept them? Nobody re-reads them. In fact, the essays you wrote are totally worthless. But the value wasn’t in the essay. What’s valuable is the effort you had to put in to produce the essay. Now, what AI essays do is they shortcut from the request to the delivery of the finished good and bypass the very part of the journey which is actually valuable—the time and effort you invest in constructing the essay in the first place.

This basically explains the whole world since about 1920. Manufactured goods, where you can enjoy extraordinary efficiencies of production—you can compress the time and effort required to make something—have massively reduced in cost. Services, which are time-dependent, have become more and more expensive. If you think there’s a hell of a lot weird with the world that was completely different when you were a kid, this is why. Like the fact that a television is almost an impulse buy, but you agonize about getting childcare. In 1920, it was the other way around. Agatha Christie had three servants in her early life but couldn’t dream of being able to afford a car.

Here are some things to think about when choosing service boundaries: Does your monolith take a long time to startup? That makes upgrades a pain, so you might want to split out the slow part to make upgrading other things go faster. Do you need the right datastore schema version? That sometimes requires lockstep upgrades/downgrades of all instances of a backend so they're on the same schema version. Lockstep upgrades are risky and tend to prevent rollbacks; you sometimes want to keep the schema-dependent part as small as possible. Are continuous integration tests frequently failing? If so, then I have bad news. Those failing tests are saying your code is broken. That's a feature! Splitting services and rolling them out separately will probably fool your tests into passing, but then you'll have compatibility and version skew problems in production instead. That's no help. Do some parts scale differently from other parts? For example, some operations are memory-heavy while others are CPU-heavy. This is not important as often as you'd think. If all your instances are load balanced properly, then the load tends to naturally spread around in a pretty efficient way. If load balancing becomes a problem, you can measure it and fix the specific granularity problem later. Do expensive requests need to be run with less parallelism? A common microservice architecture is to dump requests into a message queue and have worker instances process requests sequentially. But this goes wrong more often than you'd think, and there are better designs that avoid "queue explosion" problems. You can implement those same designs in a monolith. Do you have services with different quality/reliability targets? This can be a good reason to split services. For example, at Tailscale we have only a couple of services with very strict uptime goals: the coordination service and the log catcher service. Those two are already split for security isolation since logs are so sensitive. On top of that, our "real-time" log/metrics processing pipeline can tolerate more downtime and therefore more experimentation, so it's split from the high-reliability services and can have different deployment procedures.