The Swift community has recently had an interesting discussion about "when to use classes vs when to use structs." The conclusion I think has been most widely accepted is "use classes when you need reference semantics, and use structs when you need value semantics.".

"Classes for references, structs for values" is just a small part of a bigger picture, though. I'd argue that when you need reference semantics for something (the database, the GPU, the server), object-oriented design principles best inform how to interact with it, and when you need value semantics for something (a user record, a UI scenegraph, a transaction), functional programming principles best inform how to transform it.

The lovely thing about this is that it takes "when do I use FP vs OO?" - which used to be a very involved decision for me - and boils it down to one concrete rule of thumb:

  • If something must be modeled as a "logical singleton"¹, use OO design principles to code for it. If it's not, use FP.

This has clarified some things immensely for me, like "why does the IO monad seem like such a friction point in Haskell?" (sockets and files are singletons!) or "why does the Actor model seem to work so well for programming distributed systems in Erlang?" (each machine & message sink is a singleton). It also helped me understand why we gain so much from going 100% functional/immutable in spite of the fact that the paradigm isn't usually a perfect fit for our entire domain - there are many fewer singleton concepts in most problems than non-singleton concepts.

You can see an example of object-oriented approach sneaking into a functional language in Clojure's component state-management library. Dependencies are injected, components have constructors and destructors (the start and stop protocol methods), and the examples are singletons: databases, schedulers, and the like. Conversely, you can see a really functional approach to business logic sneaking into high-performance object-oriented systems like LMAX and seminal books like Java Concurrency in Practice.

The value/reference dichotomy also corresponds really nicely to "functional core, imperative shell", a principle for architecting programs coined and cogently analyzed by Gary Bernhardt. To paraphrase him loosely: The things at the boundary of your system (taking input from the user, drawing output on the screen, talking to the network) are part of the "imperative shell" of your program, an area in which very few decisions need to be made, but many processes need to be integrated. The computations at the core of your system (making decisions about your domain that determine what is input and output) comprise the "functional core": here your domain can be modeled immutably, unit tested in isolation, and be manipulated with pure functions. When one "imperative shell" needs to talk to another, it can do so by sending it messages that are just immutable values from the functional core - either on an explicit queue as in CSP or an implicit one as in the Actor model.

The part of your program that you use value semantics and functional programming to model (the "functional core") is well served by applying FP dogma: use immutable values, transform them with pure functions, maintain as little state as possible, and unit test.

The part of your program that you use reference semantics and object-oriented programming to model (the "imperative shell") is well served by applying OO dogma: tell, don't ask, encapsulate state, separate commands and queries, invert dependencies, minimize API surface area, and integration test.

I have a theory that value semantics are an indicator of the parts of your program that are trivially parallel, but not concurrent, and reference semantics indicate the inverse. It seems like a good sign that the parallel Haskell runtime uses purity as a criterion for work-stealing parallelism.

The final thing I'd like to note: We gain immense power from taking a concept that we only knew how to handle with reference semantics and figure out how to model it with value semantics. Think about the hype that React.js is getting now, not as a JavaScript framework, but a way to model user interface as a value - it's changing the way people think about user interface programming because of the power value semantics give us. If you've already drunk the Datomic kool-aid, you can see that it's doing the same thing - taking the database, something that used to be modeled as a reference, and enabling us to model it as a value ².

In short:

Use object-oriented programming / reference semantics / mutability / concurrent abstractions / integration testing for concepts that are best modeled as singletons.

Use functional programming / value semantics / immutability / parallel abstractions / unit testing for everything else.

¹ What do I mean by "logical singleton"? Something is a "logical singleton" if, when you make changes to it, every other part of the program can see that change - you don't have control of its scope (irrespective of whether you actually used the "Singleton Pattern"). Mutating it in one place mutates it everywhere - think writing to a file, blitting pixels to the screen, or applying a transaction to the database.

² React and Datomic hint at a general strategy for doing this: take a snapshot of the thing you have a reference to, transform the snapshot as if it's a value, then atomically sync your snapshot with the real thing. The tricky part is providing performant implementations for snapshotting & syncing back.

Huge thanks to Justin, Tim, Max, Noam, Feivel and Samer for reading drafts of this and providing feedback, to Julia for convincing me to write this up, and to Gary Bernhardt, Rich Hickey, and Rob Pike for their excellent pedagogy.