Constructors Shouldn't Throw

Posted by jcbhmr 3 days ago

Comments

This is broken. Theres no common-good in allowing invalid objects to be created. Constraints and validation are important in all parts of a system, and ctors is just one of the units that also need to perform checks to ensure this is true.

Comment by quietbritishjim 3 days ago

Absolutely agreed.

If you really wanted to go down the author's route, then you'd at least make the constructor private, and have a static method that returns a valid constructed object or an error, so you never have e.g. an "unconnected" database connection:

   class DatabaseConnection {
   public:
     static std::expected<DatabaseConnection, ConnectionError> connect(/* ... */);
     // ...
   };

But I wouldn't even bother with this - it's exceptions with extra steps. I seem to remember that Stroustrop said in the Design and Evolution of C++ that the main reason for adding exceptions in the first place was to allow a failure result from a constructor.

Comment by quuxplusone 3 days ago

> The C++ container classes assume noexcept constructors for this reason, because creating a list of size 20 calls your constructor 20 times, and the cleanup gets so much messier.

No, STL containers such as std::vector definitely do not assume (nor require) non-throwing constructors. They just do the right thing. Yes, cleanup is messier than it would otherwise be.

STL containers do require that your destructors not throw (but do not require that they be noexcept).

Comment by quietbritishjim 3 days ago

Containers also give more efficient and simpler semantics if move constructors are noexcept, which they generally should be.

Comment by quuxplusone 2 days ago

std::vector does. No other container cares about noexcept at all.

https://stackoverflow.com/questions/66459162/where-do-standa...

Background: https://quuxplusone.github.io/blog/2022/08/26/vector-pessimi...

Comment by quietbritishjim 23 hours ago

Ah yes, I forgot that only vector can relocate elements. (Other containers can invalidate iterators but the objects themselves aren't moved.)

Comment by account42 2 days ago

They could be even more efficient if we had a way to indicate relocatability instead of having to fall back to more general move semantics.

Comment by xg15 2 days ago

> The C++ container classes assume noexcept constructors for this reason, because creating a list of size 20 calls your constructor 20 times, and the cleanup gets so much messier.

I find it interesting that programming newbies are usually scared of C/C++ because "you have to deal with pointers" unlike in Java, Python, JavaScript etc.

It's interesting because all those languages of course use pointers and heap-allocated objects. The thing that sets the other languages apart is that they only allow pointer access and thus can make the language easier to reason about.

So what's "unusual" in C, C++ and Rust is really the parts that are not pointers, i.e. the ability to manage a data structure and its layout in memory directly, with all its consequences.

Comment by msm_ 2 days ago

I don't think this is the crux of the matter. Programming newbies are not scared of C/C++ because of std::unique_ptr<T>. They are scared of three-star programmers, of intrusive lists, of clever packed data structures - in short, of pointer arithmetic. Java, Python and JavaScript don't allow that, so it's hard to compare them (that's why we distinguish between "pointer" and "reference", even though in practice they are exactly the same thing under the hood, of course).

Comment by x-complexity 3 days ago

Constructors are dedicated language-sanctioned functions that transform (data inputs X) into (type/object Y).

The given X should be subjected to the requirements necessary to create a valid Y. If the requirements fail, you have two main choices: Create an invalid Y, or reject X & return an error in however you want it to be returned.

The former leaks said rejections throughout the rest of the type's/object's code, & makes more work for me to parse through where & how the problem occurs.

The latter option creates a cleaner system altogether, due to rejecting sections of invalid states. Not all will be rejected, but the fact that some do get rejected already makes its workload comparatively lighter than the former.

I know which my lazy bum would pick.

Comment by nitwit005 1 day ago

> Okay I’m not in the business of writing Date libraries, but still, if you need to parse a string, write a parser. And then write a factory method Date.parse(string) that uses that parser and calls a private constructor.

That's just a constructor you've given another name.

Comment by AnimalMuppet 3 days ago

> Because the semantics of a constructor is to bring an object into a valid state and you don’t expect it to perform any functionality.

No, I expect it to perform the functionality of bringing the object into a valid state - whatever functionality that takes. And if it can't, then throwing is the next option, because it's preferable to having an object in an invalid state.

Comment by robotresearcher 2 days ago

What if the object can be invalidated asynchronously after the constructor has completed and before you get to use the object? For example by a remote connection dropping? You have to check that it still works anyway: you can’t assume it by construction.

I prefer to have constructors that can’t go wrong (except trivially for OOM that ends the whole process), and then manage their dynamic state after construction.

Comment by AnimalMuppet 2 days ago

Well, what if a valid object has to have a semaphore, because it can be accessed by things from multiple threads, and the OS has a limited number of semaphores that it will create? The object won't be in a valid state without a valid semaphore, so the constructor can fail. (Or, if you want, the semaphore's constructor can fail, but that will mean that this object's constructor also fails.)

Look, for your problem, you probably want to do what the fine article said - to have "invalid connection" be a valid state of the object.

Comment by robotresearcher 2 days ago

I’d initialize the semaphore and any other things that could fail in a post-constructor setup/startup method. It’s a little ugly and it violates RAII, but I think it’s worth it on balance. So yes, I’m agreeing with you that’s how I’d do it.

Comment by AnimalMuppet 2 days ago

I've done it that way at least once. But, personally, I prefer RAII, rather than a "half-constructed" state.

Comment by Panzerschrek 3 days ago

> Turn your Database class into a state machine with an uninitialized and a connected state. Then have a Database.connect()

It's a bad advice. It's preferred instead to have a connect function returning a database instance in a successful case or an error otherwise.

Comment by account42 2 days ago

Is it? Your database connection can fail at any time even if initially successful so what's the advantage of having another separate error path for the initial connection.

Comment by oraro 3 days ago

Trying to keep an open mind here–if your constructor arguments are plain wrong then who should be throwing the error then? The Factory? The Builder? And what if they don't exist in the first place?

Comment by robotresearcher 2 days ago

Ideally you’d use the type system to ensure that arguments can’t be wrong.

For example, if arg values can’t be negative, use an unsigned arg type.

Comment by verdverm 3 days ago

neither should destructors, my first internship was at Bloomberg and I put a BIG into a crash loop with this

Comment by quietbritishjim 3 days ago

Ah, the "big". That's some fun insider knowledge!

At one point they needed a custom build of AIX because its executable size limit of 1GB wasn't enough for the big. This was about 20 years ago, I suppose that seems quaint now.

Comment by verdverm 3 days ago

when I was there, the bigs were 4 gig, ~2009±1

for others context, bigs were the compiled binary deployed to a server, a composition of BB terminal functions, bin packing style

Comment by DemocracyFTW2 3 days ago

> Constructors Shouldn't Throw

wrong

Comment by verdverm 3 days ago

this is not helpful, please explain why you think this

If we interpret "constructor" more generically as setup, one of my preferred languages, Go, does not have constructors, but we do have similar setup functions, which typically return errors, and we handle them as soon as possible without breaking out from the designed control flow.

Comment by DemocracyFTW2 3 days ago

I think elsewhere in this thread the point has already been made so I thought to spare the Ink. A constructor is essentially a function that returns a value. There's little reason to think that constructors are more than that. A few examples: in JS, a constructor is a method called `constructor()`. It's a bit special in that it's the method that is privileged to be the one called when the class name is preceded by the keyword `new`, but that's almost about it. It can be replaced or supported by one or more factory functions, and this is what happened with NodeJS's `Buffer` class and generals JS's `Array` class. In Ruby, the constructor is just a method called `new()`, and I guess it's not special beyond there being a convention to construct new instances with a method called `.new()`. Like any function, a constructor has an input and an output value. For example, in `new Point 3, 4` the input is the pair of numbers `( 3, 4 )` (here written spread out, but it doesn't matter), and its output is an instance of class `Point`. Say you have a class List that accepts as only parameter the length (number of elements) of the `List`. When I call a function `build_list length: -4` I expect the `length` argument to be suitably interpreted, or an error to be thrown. Most of the time a list with a negative number of elements does not make a lot of sense, so the natural thing is to indicate an error, which in many current languages is done by throwing an exception so as to prevent silent failures. When I call a constructor as `new List length: -4` I naturally expect the exact same considerations to apply.

Comment by stevekemp 3 days ago

In go there are the "init" functions, but those do not return a possible error result.

I guess you're thinking more of package constructor methods, such as NewFoo, which tend to return the new object and an error value.

Comment by verdverm 3 days ago

yes, I use init sparingly and generally panic, for low level program startup, versus the NewFoo functions, which are for dynamic objects the way constructors are used, I use NewFoo -> (Foo, error) for internal runtime, client, and configuration handling as well

Comment by 2 days ago