[swift-evolution] [Review] SE-0159: Fix Private Access Levels

[Drew Crawford]

There are a couple of problems with the design which are causing Cognitive Dissonance:

1) Private & fileprivate require very different mental models to operate them
2) They are named very similarly to each other
3) There is a third concept of “private to the type” which requires a third mental model, but is spelled the same way in other languages 

How would you respond to the allegation that supporting different "redundant" semantic models is actually a *design goal* of Swift?  For example:
Value semantics vs reference semantics
Pass-by-owned vs pass-by-shared vs pass-by-reference parameters
dynamic dispatch vs static dispatch
associatedtype vs generics
"complete" types (e.g. Array<Int>) vs  "incomplete" types (Sequence) and which one is permitted after a colon
strong vs weak vs unowned
I would argue that supporting whatever the programmer's chosen mental model is actually Swift's greatest strength.  We could have a language with only reference types for example, it would be far, far simpler and easier to teach.  I am glad that we don't have that language.  I don't use all of these in every program, but I use all of them eventually, and each one is right for some situation.

Many of us don't know half these models half as as well as we'd like.  But we also like less than half of them half as well as they deserve.  Visibility is in line with the trend here.

Of the three, I am personally most supportive of file-based


My concern about dropping down to file-based is it doesn't allow any nesting, which isn't powerful enough for several examples posted in this discussion.  Unless by "file-based" we mean "directory-based" which is another proposal that has seen strong opposition from the "it's too complicated" wing.


On March 24, 2017 at 5:00:33 AM, Jonathan Hull via swift-evolution (swift-evolution at swift.org) wrote:


Several people have asked:  What is the harm it causes?

I would like to provide a (hopefully) clear answer to that.

There are a couple of problems with the design which are causing Cognitive Dissonance:

1) Private & fileprivate require very different mental models to operate them
2) They are named very similarly to each other
3) There is a third concept of “private to the type” which requires a third mental model, but is spelled the same way in other languages 

Any one of these would cause confusion.  All of them together mean that even those of us who understand how they work are bothered by them on a subconscious level (similar to how misaligned graphics will give everyone a subconscious feeling that a graphic design is “off” even if they can’t discern why consciously).  Even many of those who are arguing to keep private have argued to rename things or to make it work in a more type-based way by extending visibility to extensions.  I don’t think anyone is denying there is a problem here with the status quo.

There is a general rule in design that things that look similar (or in this case are named similarly) need to behave similarly to avoid causing confusion (and conversely things that behave differently *need* to look different).

**This is not something that will go away simply by learning the behavior.  It will continue to be a rough spot in the language that causes discomfort until the underlying design issue is fixed.**

The ideal solution here would be to choose a single model and make everything coherent under that.  We choose either file-based, scope-based, or type-based access control and unify under that.

For file-based, that means that our sub-modules, when we add them, would need to be file-based as well.

If we choose Scope-based, then it really cries out for some sort of parameterized private.  That is you only have private and public, but for private you would be able to name a scope to widen the visibility to the named scope.  Submodules would then need to be scope based, and you would probably use the parameterized private with the submodule name to share data within it.

Type-based is very common in other languages, but I fear it would not play nicely with the way Swift has grown to work.  The pros are that you could access private variables from extensions, but getting types to work together with the ease they do now would require a lot of design work (and may not even be quite possible).  You also have additional complexities around subtypes, etc...


We really need to choose a single direction and not mix and match.

Of the three, I am personally most supportive of file-based because it is the closest to where we are now, and it seems to simplify many of the concerns about “friend” types by having visibility be dependent on proximity… which seems very intuitive to me.

As I said before, my preference would be to tentatively accept the change, but delay implementation until we design submodules for Swift 5. That way the design can be more cohesive as we can tackle it all at once.  If submodules won’t be in scope until after Swift 5, then we should implement the change now.

I hope that explanation was at least a bit helpful…

Thanks,
Jon

Technical note:  For the curious, basically we have made a design which is "locally consistent, but globally inconsistent".  That is why the effect doesn’t go away after learning the concepts involved.  Basically we have made the programming equivalent of the following image:

https://upload.wikimedia.org/wikipedia/commons/7/7f/Impossible.png
 

_______________________________________________
swift-evolution mailing list
swift-evolution at swift.org
https://lists.swift.org/mailman/listinfo/swift-evolution
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20170324/a19312f2/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: F0F862F8-0489-49F0-A06A-BEB6B7AE5C92 at hsd1.ca.comcast.net.
Type: application/octet-stream
Size: 2864 bytes
Desc: not available
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20170324/a19312f2/attachment.obj>