SE-0117 (original, revised):
The major observation here is that not all classes make sense to subclass, and
it takes real thought and design work to make a class subclassable well. As
such, being able to subclass a public class should be an additional “promise”
beyond the class just being marked public. For example, one must consider the
extension points that can be meaningfully overriden, and document the class
invariants that need to be kept by any subclasses.
Beyond high level application and library design issues, the Swift 1 approach is
also problematic for performance. It is commonly the case that many
properties of a class are marked public, but doing this means that the
compiler has to generate dynamic dispatch code for each property access. This
is an unnecessary performance loss in the case when the property was never
intended to be overridable, because accesses within the module cannot be
devirtualized.
Chris Lattner:
As expected, this proposal was extremely polarizing, with valid arguments on both sides. The opinions held by supporters and opposers are held very strongly, and hundreds of emails were generated in a healthy debate about this topic.
[…]
On the first point, there are three related arguments against SE-0117:
First is that clients of Apple frameworks often choose to subclass classes that Apple publicly documents as being “not for subclassing”, as a way of “getting their job done,” typically as a way to work around claimed bugs in Apple frameworks. The core team and others at Apple feel that this argument is analogous to the argument that Swift should “support method swizzling by default”. Swift loves dynamic features, but has already taken a stance against unanticipated method swizzling, by requiring an API author to indicate where they allow method swizzling with the ‘dynamic’ keyword.
Yes, it’s absolutely analogous. It’s become obvious that “unanticipated” dynamism is not something that Apple wants to support with Swift. It’s clearly not Swifty, if by Swifty you mean the vision of the core team.
Second is that clients of some other public API vended by a non-Apple framework (e.g. a SwiftPM package) may end up in a situation where the framework author didn’t consider subclass-ability, but the client desires it. In this situation, the core team feels that a bigger problem happened: the vendor of the framework did not completely consider the use cases of the framework. This might have happened due to the framework not using sufficient black box unit testing, a failure of the imagination of the designer in terms of use cases, or because they have a bug in their framework that needs unanticipated subclass-ability in order to “get a job done”. Similar to the first point, the core team feels that the language is not the right place to solve this problem. Instead, there is a simple and general solution: communicate with the framework author and get them to add the capabilities that you desire.
Yes, there’s a bigger problem, and, yes, the language can’t “solve” this problem. That doesn’t imply that the language should forbid building a temporary workaround. In my view, the language should be helping to get the job done, not enforcing an idealized solution that may not even be possible because the vendor may be busy, uncooperative, or no longer available.
Third is a longer-term meta-concern, wherein a few people are concerned that future pure-Swift APIs will not consider subclass-ability in their design and will accidentally choose-by-omission to prevent subclass-ability on a future pure-Swift API (vended by Apple or otherwise). The core team feels that this is an extremely unlikely situation for several reasons. First of which is that it heavily overlaps the first two concerns. More significantly, any newly-designed and from-scratch APIs that are intended for Swift-only clients will make use of a breadth of abstractions supported by Swift—structs, enums, protocols, classes.
Indeed, this proposal is just one piece of a large trend. Swift is chipping away at dynamism for classes: subclassing, swizzling, hiding private classes and private methods. And the future, it appears, is APIs that rely more on structs, enums, and protocols, which are even more static. There are definitely some benefits to this direction, but I am concerned that it will lower app quality over the long term. This specific proposal may not have a huge impact because it affects neither Objective-C APIs (past) nor non-class Swift APIs (future).
To reiterate, as a summary, the core team agrees with conviction that it is the right default for public classes to be non-subclassable outside their module, unless they carry some additional indication by the API author that the class was designed to be subclassed.
Here’s my slightly exaggerated summary of the two points of view:
Me: I agree that dynamism shouldn’t be abused, but there are library bugs and limitations, and sometimes we need these tools to cope with them. I hate coding these workarounds and would love to remove them and unclutter my code as soon as the underlying issues are resolved. But bugs that affect the customer are much worse than temporary ugly code. Of course, ideally library vendors wouldn’t ship any bugs, but this is the real world. Nudging the API author to “think carefully” is well-intentioned but naive. The proposed solution is worse than the problem.
Apple: Making library vendors opt into dynamism will encourage them to think more about their designs. They will foresee how their APIs will be used. Making it impossible for developers to work around library bugs will foster communication with the vendor, who will then fix them. The fixing will be easier because the vendor won’t have to take into account developers’ workarounds. We don’t particularly care if your app is broken for months or years while waiting for a bug to be fixed. Developers won’t feel pressured to work around library bugs because competing apps will be subject to the same restrictions.
Here are some of the more interesting comments from the discussion:
Kevin Lundberg:
I do not want to be constrained by authors of libraries or
frameworks into interacting with a system in only the ways they forsee.
By making the default be non-subclassable, if a designer does not put
thought into all the ways a class can be used then I as a consumer of
the library am penalized.
Another point I can think of is mocking in tests. I’m unaware of any
fully featured mocking solution in swift, and subclassing a class to
override methods for mocking purposes is one way to make unit tests work
with third party dependencies. If a class was not designed for
subclassing, then writing tests where the class’s behavior needs to be
suppressed would not be possible.
[…]
I’ve had issues before when working with third party C# code where the
author didn’t declare a method as virtual, but the problem I was trying
to solve required me to override that method. Luckily the code was open
source, so I just embedded and edited the source into my project instead
of using the prebuilt binaries, but that becomes a maintenance headache,
and it is not possible for components that do not make the source available.
L. Mihalkovic:
Can’t really help for feel like it is training wheels all around… or padlocks on every kitchen cupboards.
Kevin Lundberg:
It’s not possible to
accidentally override a method like it is in java or objective-c, so the
proposal won’t help people who might accidentally override something. If
a developer tries to override a method or class they probably have a
specific reason in mind for doing so, and taking away the options that
this ability presents, by nature of it being the default behavior,
doesn’t sit well with me as it has the potential to take away a good
tool from my toolbox.
I agree that API authors should have the power to restrict this
dimension of a type’s usage, but I feel that it should be a conscious
choice to do so.
Karl:
I really enjoy designing elegant, well-defined APIs and this behaviour for classes has bothered me for some time.
On the other hand, I have also subclassed things that weren’t meant to be subclassed (e.g. I believe UITabBar was an awkward one back in the day, and UINavigationController has so many broken behaviours it’s almost a requirement to subclass and patch it, even though I believe Apple disapproves). I could fall back to Objective-C, but that’s an implementation detail. In some theoretical future with an all-Swift UIKit, what do I do about those issues?
After thinking about it, I decided that in this theoretical future, no App would be able to use those hacks, and so all Apps would have the same broken behaviours and degrade the quality of the platform to such an extent that Apple is forced to do something about it. Ultimately, that’s exactly what we want; they need to see broken Apps everywhere in order to prioritise fixes. That improves code quality all-around.
Tino Heth:
I can’t fight the feeling that many fans of ideas like this believe that good designed libraries come for free by simply adding restrictions — which, at least in my opinion, just isn’t true:
No matter what the defaults are, good libraries are hard to build, so I predict this proposal would not only fail in increasing framework quality, but also will make it much harder for users of those frameworks to work around their flaws, which are just a natural part of every software.
Michael Peternell:
I think it’s a step backwards. […] It’s not bad to allow sealing of classes. I don’t see real value in it though. And “sealed” should not be the default. Either the class if final or not: for me that’s a valuable distinction. “sealed vs. unsealed” is not.
Paul Cantrell:
In all the OO languages I’ve used, classes are open by default. That raises a question: is this really a good idea? Aren’t we ignoring established precedent? What about all the unauthorized subclassing we’ve done in the past?
I’m sympathetic to those accustomed to Objective-C, Ruby, Javascript, Python, and other languages with more fungible type systems who argue that it’s unreasonable to expect us all to use classes only as library authors intend. Monkey patching has saved many a day, they’d say. And it’s true, it has! I rely on it.
My counterargument: the ecosystem is changing.
[…]
The big exception to this is Apple itself, and the direction of this proposal implies a large cultural shift in how Apple deals with its developer community. Having the Swift language so aggressively prevent the dubious, brittle, bad-idea-but-it-gets-the-job-done workarounds that the Obj-C runtime allowed means that Apple’s platforms are going to have to be more consciously extensible, more transparent in their design decisions, and more responsive to unanticipated needs.
Tino Heth:
With “sealed by default”, the situation changes: Users are protected from some simple bugs, but nonetheless, they’ll encounter situations where the library doesn’t do exactly what they want. So, you take a look at the source, find the problem, and fix it. It’s no bug at all, it’s just a tiny degree of freedom that is missing because it wasn’t important to the author. You can create a pull request as well, but it doesn’t offer a real improvement to the author, who is already busy with dozens of similar requests by other users -> you end up with a custom branch with all the housekeeping associated with it.
So overall, I’m quite sure that the proposal won’t improve software quality, but rather degrade it.
Ricardo Parada:
I am afraid that developers will not take the time to specify which methods are overridable resulting in libraries that are difficult to patch, extend.
In my 26+ years of object-oriented design and programming (other languages, Objective-C since 1990 and Java since 2001) I have worked with object oriented libraries and subclassed methods that the authors probably never anticipated. I have been able to fix problems, enhance classes by creating subclasses with fixes and enhanced behavior.
In java for example I have seen that sometimes I would have been able to fix bugs or enhance the existing classes had the author not chosen a method to be protected or private. Sometimes they had a good reason but sometimes they didn’t. Is have been able to survive using an awesome library that was discontinued and end-of-lifed thanks to subclassing that has allowed me to fix problems and enhance over the years as the Java language kept evolving.
Jakub Suder:
I agree that subclassing library classes that were not explicitly planned
to be subclassed might sometimes lead to bugs. But just because a technique
or feature of a language allows you to sometimes make mistakes, it doesn’t
mean it should be completely disallowed. By this logic, we’d have to remove
all instances of "!" any anything related to pointers from Swift. It’s
possible for a feature to allow you to shoot yourself in the foot, and
still be useful at the same time.
[…]
I don’t think the problem is significant, and I’m afraid the proposed
solution is worse than the problem itself.
Jonathan Hull:
I don’t think this proposal will achieve its stated objective of forcing people to think about subclassing more. It will just add confusing boilerplate.
Things like Swift optionals work well because they make the (often forgotten) choices explicit in the context that they are used. In the world of Human Factors, we call it a forcing function. This proposal has the inverse structure, and will be ineffective, because the “forcing” part of it shows up in a different context (i.e. trying to use a framework) than the decision is being made in (writing the framework). This type of thinking leads to things like Java and the DMV.
[…]
Those who were prone to be thoughtful about their design would have been anyway. Those who are not thoughtful about their design will just leave these annotations off… leaving us with no recourse to extend/modify classes. When people complain, they will add the annotations without actually thinking about the meaning (i.e. stack overflow / the fixit tells me I need to add this word to make the compiler happy). All this does is put framework users at the mercy of the framework writers.
Brad Hilton:
I […] give a strong, strong, strong -1 to this proposal. To make classes non-subclassable by default is only going to lead to unanticipated pain and frustration. Also agree with other comments that subclassable and overridable conflate access control with class behavior. If we want to make it possible to define a class as non-subclassable to external users, I’d agree to something more consistent with existing swift access control like public(final) as has been proposed by other commenters. However, I agree that making classes final by default is a bad idea that will create a larger problem that it solves.
Jean-Daniel Dupas:
I can’t count the number of times it save my hours [to be] able to override arbitrary classes and methods.
Sometimes to simply add log point to understand how the API work. Other times to workaround bugs in the library. Or even to extends the library in a way that the author did not intent in the first place, but that was perfectly supported anyway.
I already see how libraries author will react to that new default. They will either don’t care and mark all classes as subclassable, or find to burdensome to get subclassability right and prohibit subclassing all classes.
Jean-Daniel Dupas:
Nonetheless, being able to subclass any class allowed me to solve a bunch of very bad memory leaks in the new NSCollectionView implementation, and made it usable for my project.
Thinking than you will have to work only with perfectly design libraries without any bug is utopian.
And thinking you can predict every usages of your library by users is short sighted IMHO.
John McCall:
I sympathize with the argument about wanting to fix bugs and add features via override, but that’s never been maintainable in the long term; you always just end up with superclasses that everyone is terrified to touch because every subclass has its own invasive “fixes”, and that’s even when working within a single codebase. With libraries, you can pretty quickly get locked in to a specific version because your customizations don’t work with new releases; either that, or the maintainer just decides that they can’t fix of their mistakes and so goes off to rewrite it from scratch. Either way, it’s not good for the ecosystem.
Jeremy Pereira:
Thirdly, this is a huge breaking change. Not only is it a breaking change, but for a lot of people, the breakages will be outside of their control. Consider if I publish a module with a class with public methods and you subclass it in your code. Once this change is implemented, my code will still compile and pass its unit tests but your code is now broken and you are dependent on me changing my code to fix your code.
Tino Heth:
Defaults matter, because they transmit a message:
Every rule and obstacle we add to Swift is a statement that says “we favor bureaucracy over freedom”, and this will affect the community evolving around the language.
When you use a library in a way that wasn’t anticipated by its author, you’ll ran into issues occasionally; nonetheless, I think we should struggle for an open ecosystem that encourages others to experiment and fail, rather than to limit their possibilities in a futile attempt to protect them.
Garth Snyder:
I’ll just note that this proposal reminds me of the fortune file entry, “Whenever you see a sign that says ‘no exit’, it means there is an exit there.”
Specifically, under this proposal, whenever you see a nonsubclassable, non-final class (which I suppose would in fact be any class not explicitly marked as subclassable), it would mean that the library implementor IS in fact subclassing it. Otherwise, it would just be final. Therefore, it’s by definition a perfectly reasonable class to specialize. The implementor just doesn’t want YOU doing it.
Maybe the implementor has a very good reason for this. But in the modal case, probably not.
[…]
There should absolutely be a way to say “don’t subclass this.” But do you really need anything more than a comment that says “don’t subclass this?” Perhaps as a compromise, and as a concession to the value of automated checking, this feature could be reformulated as a warning system. A class marked as nonsubclassable could be extended only be including a corresponding acknowledgment keyword in the extending class, a kind of Swift version of -IHaveBeenWarnedThatAPFSIsPreReleaseAndThatIMayLoseData.
Paul Norton:
Because of that experience, I am fully convinced that library authors will never imagine all the ways developers will use a library’s API.
[…]
I also think it would eventually force a second change; some sort of override to the default, at which point we’ve extended the language further to get closer to where we started.
Andre:
Personally, Im not against sealed by default, but I think there are cases where closed source libraries have certain cases where workarounds are necessary, and just sealing by default will prevent those cases.
One could say, “well just use an open source one, or change vendors” but its not that easy in The Real World™ where we get certain SDKs shoved down our throats by the suits… and while that may be a separate issue to the one at hand, its still a problem that won’t resolve itself by simply locking down things…
In my own case, Ive fought with NSBrowser / NSTreeController in the past and the only way to resolve things was to subclass (and no, waiting 1 or 2 years for a fix is not acceptable if you already have a product in the wild).
Jonathan Hull:
Please stop saying that this proposal will bring more consideration to the design of libraries. It isn’t true. I haven’t even seen an argument for why it would be true, it is just taken for granted that it is true.
[…]
It sounds good, but at the end of the day, people are human and they will make mistakes. Best to either catch those mistakes in the context where they happen or to mitigate the effect of it. This proposal basically forces you to feel the full effect of other people’s mistakes (thinking that it will discourage them from making them in the first place).
[…]
The current proposal will only cause massive problems down the line, IMHO. We will find an escape hatch is needed, but we will have made optimizations based on assumptions of finality which prevent us from easily adding one.
L. Mihalkovic:
There are IMO no advantages to sealing by default. If there were, I cannot imagine how they can have so consistently eluded the designers and maintainers of so many great OOD languages in the last 30 years. Does it mean it is just a matter of time for the core team to take it the c++ standardization committee to make sure C++ gets enhanced the same way?
L. Mihalkovic:
After 30 years of Objc’s msg dispatching being touted as not a performance problem, all we hear for the last couple WWDC (did you pay attention to the ‘they are getting it’ during one of the session) is how much every single dynamic dispatch must be chased out of our systems, culminating with this ‘seal by default is best for you’. Why can’t developers [writing] code be made responsible for writing well or not?
Jonathan Hull:
There is also a dangerous difference between helping the programmer catch mistakes (e.g. don’t accidentally subclass the wrong method) and trying to prevent them from coding in a style you disagree with. I have been seeing far to many proposals of the second variety of late.
let var go:
It is the opposite of nearly every other OOP language’s behavior and the
opposite of what anyone will expect who is coming to Swift after doing any
OOP programming anywhere else. While I believe that Swift should be free to
buck the trends of other languages where there are significant advantages,
changing something this fundamental will introduce a new learning curve for
even experienced programmers. The advantages better be significant,
otherwise you are introducing a frustrating quirk into the language without
any great benefit in exchange. But the advantages of default
non-subclassability are marginal at best.
[…]
The developers who
don’t make careful design decisions will just go with the default - in this
case they will just leave the default in place, and their public classes
will not be subclassable. That doesn’t improve anything, it just makes
sloppy, poorly written code harder to fix in the off-chance that you are
stuck working with it. In other words, quality will not improve, but
productivity will suffer, because it will be harder to develop workarounds
for the problems that will inevitably appear in even the best-designed APIs.
[…]
But this whole issue of default non-subclassability doesn’t fall
into that category of “safety” at all. It doesn’t help a programmer produce
safer code. The same hidden flaws will persist whether non-subclassability
is the default or not. The difference is that those same hidden flaws will
be more difficult to deal with after the fact.
Colin Cornaby:
The “final should be default because adding final after the fact is destructive” arguments are interesting, but ultimately not convincing to me. If you thought your class was safe for subclassing, but it ultimately wasn’t, this solves nothing. You’ll mark your class as subclassable, and you will ship it, and there will be issues, but it will still be too late to take things back. If you know your class is not safe for subclassing, you should mark it as final. There is no advantage here in that scenario.
This is all part of building a safe public API. Public API design can be difficult, but if you don’t understand safe subclassing designs, you are likely to miss the issues and mark your class as subclassable and ship it anyway. Again, the best way to tackle this is to find better ways to surface the issues. Making final the default still doesn’t solve the core issues of people not understanding the right design.
Jonathan Hull:
My point is that you are completely ignoring an entire class of risk that has a real-world $$$ cost. Every time I have to use a framework under this proposal, I am now completely at the mercy of the author. In the case of open source frameworks I can at least make a fork, but for closed source frameworks (often from large companies where us using their framework has been negotiated by the bosses) you have now just added weeks to my development cycle while I wait for big-company-who-doesn’t-really-care-that-much to update their stuff.
[…]
Are you offering to explain to my boss/client why I can’t add the feature in a reasonable timeframe like I can with Objective C frameworks? That it may not even be possible now in Swift even though the Android guy just did it in a few hours?
Do you know what I am going to tell my boss/client? “Don’t use Swift for frameworks” and “Try to avoid partnering with companies that have Swift frameworks”. “It is too much of a risk”. “We are giving them too much control over the future of our product…” I mean, it even affects the prices that companies can charge for crappy frameworks. If I am asking for a bunch of features that I need them to add to provide a basic feature, that affects negotiations/price (vs a world where I can add it myself if needed). Sealed-by-default gives them leverage.
Tony Allevato:
The argument
shouldn’t be “we need open subclassing or we can’t fix broken APIs”; that’s
a false choice. It should be “we need something to fix broken APIs”, and
that “something” should be proposed and the need for it should be argued in
its own right.
David Sweeris:
-1 for this proposal, but +1 for solving the issues it raises
Regardless of what ends up being the defaults, I’m a very strong -1 on conflating visibility and subclassability/extendability.
Jon Akhtar:
I completely agree, the programmer should have as much power as we can
give them, even if it means allowing them to shoot themselves in the foot,
because at the end of the day Swift isn’t an academic exercise, it is a
real world programming language, so it should be optimized for solving
real world problems not having some kind of technical, philosophical, and
stylistic perfection when those come at the cost of the former.
Update (2016-07-17): Daniel Jalkut:
It’s absurd to imagine you can predict all valid subclassing patterns. […] Apple, nor any framework provider, can provide functionality sufficient to cover the breadth of innovation. Clever hacks built the industry.
To be clear, although the proposal was “returned for revision,” it is essentially accepted. It’s just the syntax that’s being debated now (e.g. changing subclassable to open).