Let it throw, Let it throw!



Today’s article will be about handling errors in Swift.
Because let’s be honest, that makes a fun post title for the season ❄️☃️.

Objective-C and the NSError of its ways

Remember Objective-C? Back then1, the main and official way for a method to tell that something went wrong was to take a NSError* by reference.

NSError* error;
BOOL ok = [string writeToFile:path
                   atomically:YES
                     encoding:NSUTF8StringEncoding
                        error:&error];
if (!ok) {
  NSLog(@"An error happend: %@", error);
}

Man that was painful. To the point that many people were tempted to not even bother to check for errors and simply pass NULL there. Not very responsible and safe.

Do you want to build a throw, man?

Frozen-Illustration-1

With Swift 2.0, Apple decided to introduce a different way to handle errors: using throw 2.

The use of throw is actually quite simple:

  • If you want to create a function that might fail, mark it with throws in its signature;
  • Inside the function itself you can use throw someError if needed;
  • At call site, you’ll have to explicitly use try in front of your calls to methods which can throw3;
  • To catch errors and handle them, use do { … } catch { … } constructs.

This looks like this:

// Define a function that can throw…
func someFunctionWhichCanFail(param: Int) throws -> String {
  ...
  if (param > 0) {
    return "somestring"
  }
  else {
    throw NSError(domain: "MyDomain", code: 500, userInfo: nil)
  }
}

// … then call it
do {
  let result: String = try someFunctionWhichCanFail(-2)
  print("success! \(result)")
}
catch {
  print("Oops: \(error)")
}

The failure never bothered me anyway

Frozen-Illustration-2

You can see that someFunctionWhichCanFail returns a plain String, which is the type returned when everything was ok. It makes it easy to call the function “normally”, first thinking (in the do { … } block) about the happy path, to handle the case where nothing wrong happens.

The only reminder that those methods can fail is the try keyword that the compiler enforces you to add in front of that call, otherwise it’s like a non-throwing function call. And then, you only write the code to handle errors in a separate place (inside the catch)

Note that you can write more than one line (and try-call more than one throwing function) in that do block. If everything is successful, it will execute them as expected, but as soon as one of them fails it will jump out of the do block into the catch statement instead. That’s very handy as well for long runs of code with multiple potential points of failure, as you can handle them all in a single error path at the end.

NSError is a bit of a fixer-upper

Frozen-Illustration-3

Ok, but with this example we still have to handle errors using NSError, which is a pain. Comparing domains and error codes with == and make a list of domain and code constants, just to know which error we got and handle it properly… ouch.

But we can fix this fixer-upper… up with a little bit of love! What if we used what we learned in my Enums as Constants article and use enums to represent errors instead?

Well, Good news, everyone!™, that’s exactly what Apple intended in this new error handling model! In fact, when a function throws, it can throw any object which conforms to ErrorType. NSError is one of those types, but you can make your own, and it’s even recommended!

The best fit for an ErrorType type is an enum, which can even have associated values if needs be. For example:

enum KristoffError : ErrorType {
  case ClumsyWayHeWalks
  case GrumpyWayHeTalks
  case PearShapedSquareShapedWeirdnessOfHisFeet
  case NotWashedSince(days: Int)
}

Then you can now use throw KristoffError.NotWashedSince(days: 3) in a function to throws this kind of error, then use catch KristoffError.NotWashedSince(let days) at call site to catch such errors:

func loveKristoff() throws -> Void {
  guard daysSinceLastShower == 0 else {
    throw KristoffError.NotWashedSince(days: daysSinceLastShower)
  }
  ...  
}

do {
  try loveKristoff()
}
catch KristoffError.NotWashedSince(let days) {
  print("Ewww, he hasn't had a shower since \(days) days!")
}
catch {
  // Any other kind of error, whatever it is
  print("I prefer we stay friends")
}

That’s way easier to catch specific errors than having to compare domains and error codes!

That also makes better errors with clear names as constants and associated values. No more obscure userInfo, you have a clear list of associated values right there in the enum case of the error — like days in the above example — which are only valid for that specific type (wouldn’t make sense to have that days associated value for the ClumsyWayHeWalks error).

I can’t hold it back anymore

Frozen-Illustration-4

When you call a throw-ing function, the error it throws can be caught in the calling function using a do…catch. But if it isn’t, then it propagates to the upper level:

func doFail() throws -> Void { throw  }

func test() {
  do {
    try doTheActualCall()
  } catch {
    print("Oops")
  }
}
func doTheActualCall() throws {
  try doFail()
}

Here when doFail is called, the potential error is not caught by doTheActualCall (there is no do…catch capturing it), so it propagates up to the calling test() function. Because it doesn’t catch all errors, doTheActualCall must also be marked as throws: even if it doesn’t throw errors by itself, it can still propagate some. It can’t keep the error to itself, it has to… let it go to the upper level.

On the other end, test() catches all errors internally so even if it calls a throwing function (try doTheActualCall()), all errors thrown by that function are caught in the do…catch block. The test() function itself doesn’t throw, so callers don’t need to know about this internal behavior.

Conceal, don’t feel, don’t let them know

Frozen-Illustration-5

You may have wondered by now how to know which kind of error each method throws. Indeed, functions are marked with throws but what ErrorType can this function actually throw? Can it throw KristoffErrors, JSONErrors, other? Which ones do I need to catch?

Well that’s a problem. Currently, due to some ABI and resilience concerns4, this is not possible. The only way to know is using the documentation of your code.

But that’s also a good thing. For example, imagine you use two libraries, MyLibA containing a function funcA that throws errors of type MyLibAError, and MyLibB with a function funcB that throws errors of type MyLibBError.

Then you want to create your own library MyLibC, a wrapper around those two libraries, with a function funcC() calling both MyLibA.funcA() and MyLibB.funcB(). So the resulting function funcC might throw either errors of type MyLibAError or MyLibBError. And if you add another level of abstraction, it gets worse, with more and more error types being able to be thrown. If we had to list them all, and the call site would need to catch them all, that would make a quite verbose signature and catch code.

Don’t let them in, don’t let them see

Frozen-Illustration-5

For that reason, but also to prevent your internal errors from bleeding across your library boundaries and to limit the number of error types that must be handled by your users, I suggest that you keep your error types scoped to each level of abstraction.

In the example above, instead of making funcC directly propagate both MyLibAErrors and MyLibBErrors, you should instead throw MyLibCErrors. I suggest this for two reasons, both related to abstraction:

  1. Your users shouldn’t have to know which internal library you’re using. If some day in the future you decide to switch your implementation to use SomeOtherPopularLibA instead of MyLibA, which will obviously not throw exactly the same errors, the caller of your own MyLibC framework shouldn’t need to know or care. That’s what abstraction is all about.
  2. The caller shouldn’t have to handle all the errors. Surely you can catch some of those errors and consider them internal: not all errors thrown by MyLibA make sense to be exposed publicly to your users, for example a FrameworkConfigurationError error mentioning that you misused the MyLibA framework and forgot to call its setup() method or whatever should not make its way to the user, as the user can’t do much about it. That kind of error is your fault, not theirs.

So instead, your funcC should probably catch all MyLibAErrors and MyLibBErrors, wrap them / re-interpret them and expose them as MyLibCErrors instead. That way, the users of your framework don’t have to know what you’re using under the hood. You can switch your internal implementation and libs used at any time, and you expose to the user only the errors they might care for.

We finish each others sandwiches5

Frozen-Illustration-6

There is a lot more to tell about throw and the Swift 2.0 error model. I could talk about try? and try!, about the rethrows keyword for high-order functions.

I won’t have time to talk about every subject about error handling there, that would have made that post way too long; but other interesting blog posts might help you finish your exploration in the world of Swift error handling, including (but not limited to) those:

Let me finish this article by wishing you all happy holidays ☃️❄️ and see you soon for the next post!

Happy Holidays

  1. For more info about how the old way to handle errors in Objective-C work, see this great article by NSHipster. But today’s article is about Swift and the new way, so let’s not lose too much time about that old beast. 

  2. Despite its name, throw is not about throwing exceptions like in Java or C++ or even ObjC. But the way to use it is so similar that Apple decided to keep the same wording so that people used to exceptions find it natural. 

  3. This is enforced by the compiler, the aim being to let you realize that this function can go wrong and you have to handle the potential error 

  4. Swift 2.0 doesn’t support typed throws, but there is a discussion about adding that feature in the swift-evolution Mailing List where Chris Lattner explains why this was not possible in Swift 2 and why we need the resilience model of Swift 3.0 to be able to make that feature consistent. 

  5. Ok, promise, that was my last shameful Frozen reference.