Say I have a web application with UserController. Client sends a HTTP POST request that is about to be handled by the controller. That however first must parse the provided json to UserDTO. For this reason there exist a UserDTOConverter with a method toDTO(json): User.
Given I value functional programming practices for its benefits of referential transparency and pure function the question is. What is the best approach to deal with a possibly inparsable json? First option would be to throw an exception and have it handled in global error handler. Invalid json means that something went terrible wrong (eg hacker) and this error is unrecoverable, hence the exception is on point (even assuming FP). The second option would be to return Maybe<User> instead of User. Then in the controller we can based on the return type return HTTP success response or failure response. Ultimately both approaches results in the same failure/success response, which one is preferable though?
Another example. Say I have a web application that needs to retrieve some data from remote repository UserRepository. From a UserController the repository is called getUser(userId): User. Again, what is the best way to handle the error of possible non existent user under provided id? Instead of returning User I can again return Maybe<User>. Then in controller this result can be handled by eg returning "204 No Content". Or I could throw an exception. The code stays referentially transparent as again I am letting the exception to bubble all the way up to global error handler (no try catch blocks).
Whereas in the first example I would lean more towards throwing an exception in the latter one I would prefer returning a Maybe. Exceptions result in cleaner code as the codebase is not cluttered with ubiquitous Eithers, Maybes, empty collections, etc. However, returning these kinds of data structure ensure explicitness of the calls, and imo results in better discoverability of the error.
Is there a place for exceptions in functional programming? What is the biggest pitfall of using exceptions over returning Maybes or Eithers? Does it make sense to be throwing exceptions in FP based app? If so is there a rule of thumb for that?
CodePudding user response：
You're asking about a couple of different scenarios, and I'll try to address each one.
Input
The first question pertains to converting a UserDTO (or, in general, any input) into a stronger representation (User). Such a conversion is usually self-contained (has no external dependencies) so can be implemented as a pure function. The best way to view such a function is as a parser.
Usually, parsers will return Either values (AKA Result), such as Either<Error, User>. The Either monad is, however, short-circuiting, meaning that if there's more than one problem with the input, only the first problem will be reported as an error.
When validating input, you often want to collect and return a list of all problems, so that the client can fix all problems and try again. A monad can't do that, but an applicative functor can. In general, I believe that validation is a solved problem.
Thus, you'll need to model validation as a type that isomomorphic to Either, but has different applicative functor behaviour, and no monad interface. The above links already show some examples, but here's a realistic C# example: An applicative reservation validation example in C#.
Data access
Data access is different, because you'd expect the data to already be valid. Reading from a data store can, however, 'go wrong' for two different reasons:
- The data is not there
- The data store is unreachable
The first issue (querying for missing data) can happen for various reasons, and it's usually appropriate to plan for that. Thus, a database query for a user should return Maybe<User>, indicating to the client that it should be ready to handle both cases: the user is there, or the user is not there.
The other issue is that the data store may sometimes be unreachable. This can be caused by a network partition, or if the database server is experiencing problems. In such cases, there's usually not much client code can do about it, so I usually don't bother explicitly modelling those scenarios. In other words, I'd let the implementation throw an exception, and the client code would typically not catch it (other than to log it).
In short, only throw exceptions that are unlikely to be handled. Use sum types for expected errors.
CodePudding user response：
TL;DR
If there are Maybes/Eithers all over the codebase, you generally have a problem with I/O being mixed promiscuously with business logic. This doesn't get any better if you replace them with exceptions (or vice versa).
Mark Seemann has already given a good answer, but I'd like to address one specific bit:
Exceptions result in cleaner code as the codebase is not cluttered with ubiquitous Eithers, Maybes, empty collections, etc.
This isn't necessarily true. Either part.
Problem with Exceptions
The problem with exceptions is that they circumvent the normal control flow, which can make the code difficult to reason about. This seems so obvious as to barely be worthy of mention, until you end up with an error thrown 20 calls deep in a call stack where it isn't clear what triggered the error in the first place: even though the stack trace might point you to the exact line in the code you might have a very hard time figuring out the application state that caused the error to happen. The fact that you can be undisciplined about state transitions in an imperative/procedural program is of course the whole thing that FP is trying to fix.
Maybe, Maybe not: It might be Either one
You shouldn't have ubiquitous Maybes/Eithers all over the codebase, and for the same reason that you shouldn't be throwing exceptions willy-nilly all over the codebase: it complicates the code too much. You should have files that are entry points to the system, and those I/O-concerned files will be full of Maybes/Eithers, but they should then delegate to normal functions that either get lifted or dispatched to through some other mechanism depending on language (you don't specify the language). At the very least languages with option types almost always support first-class functions, you can always use a callback.
It's kind of like testability as a proxy for code quality: if your code is hard to test it probably has structural problems. If your codebase is full of Maybes/Eithers in every file it probably has structural problems.