How small should your microservice be?

, Nov 17, 2014

Given that microservices are supposed to be, well, “micro”, there’s a lot of discussion about the right size. A typical answer to this question is: A microservice should do just one thing. I don’t really think that’s an answer, as “one thing” leaves a lot of room for interpretation. But I’ve seen people suggest that each individual microservice should be as small as a single function, and I strongly disagree with this for almost every situation. Consider, for example, a function that computes something based on three input values, and returns a result. Is that a good candidate for a microservice, i.e. should it be a separately deployable application of its own?

I believe it’s easier to approach this from the opposite direction. Let’s take an example: A web-based email system. Let’s not overcomplicate things and assume it’s traditional and offers the minimal features you’d expect, such as logging in and out, maintaining some user settings, creating messages (from scratch or by replying to or forwarding an existing one), deleting messages, viewing your inbox, moving messages into folders (that you can create, view and delete), maintaining an address book, search for messages … I’m sure you get the picture. At one extreme, we could absolutely build this as a single application and ensure it’s built not as a single package, but using a reasonable internal modularization strategy. We could decide to write its core as a set of collaborating classes, maybe adhering to the DDD approach (which would classify the classes according to the role they play). Then we’d add the dependencies to the outside world, such as the UI, the data storage, external systems (such as maybe external mail handlers, LDAP directories, etc.), possibly using a layered or hexagonal architecture.

The team(s) working on this application would need to synchronize very tightly, as the whole thing is released at once. It will also be scalad in an all-or-nothing fashion, and it will be down or up and running completely, not partially. That may be perfectly fine! But let’s just assume (again) you deem it’s not, and want to cut it apart into separate parts that have their own life-cycle and are isolated from each other.

How would you go about decomposing this into separate applications or services? First of all, the login/logout stuff (the auth system) is a good candidate, as is the user profile. They could go into one service, but if we consider that the auth system has to maintain passwords (or rather, password hashes), it makes sense in my view to treat it differently from the rest. The emails and folders themselves seem quite cohesive to me, though: You could separate them, but I probably wouldn’t. If there are multiple ways to connect to the outside world, say, via the Web interface, POP3, IMAP, and SMTP, I can imagine each of those being their own service. Maybe I’d factor out the storage of messages into its own service, one that doesn’t know the difference between a document and an email. I think the address book, including its data storage, its UI and its API seems like a natural candidate to be separated from the rest.

But in all, I’d probably end up with a dozen, maybe twenty or thirty services (or self-contained systems, as I prefer to call them). And more importantly, I think that for any given interaction triggered by some outside event – like e.g. a user clicking a button after entering data into a form – I’d end up touching maybe 3-5 of them.

In other words, I think it’s not a goal to make your services as small as possible. Doing so would mean you view the separation into individual, stand-alone services as your only structuring mechanism, while it should be only one of many.