WWDC is just around the corner, and most of us will be fully focussed on what happens with SwiftUI. The unveiling last year generated great excitement, but there are still plenty of holes to be filled before it is a solid replacement for UIKit and AppKit (…or even a solid complement).
One of the big questions on my mind is how well the fully declarative approach scales to complex apps. You can already build quite reasonable portal apps for your favorite web service with SwiftUI, which is 90% of the iOS app market, but I am much more interested in the other 10%. Can you build an advanced iOS app or serious macOS app with SwiftUI? Could you develop an app like Keynote or Xcode using SwiftUI?
I spent a lot of time this year — probably more than desirable — working with the framework. At Agenda, we decided to ‘kick the tyres’ by adopting it in a new internal tool that we wanted to develop. The tool was not trivial either: it loaded significant amounts of data, shared it between multiple users with LLVS backed by the CloudKit public store, and allowed collaborative editing and discussion. Effectively, it was like a fully decentralized web app, built with native UI.
Ambitious? Sure. Too ambitious? Undoubtedly. But we learnt a lot before abandoning the idea. And it did actually work, though after a while we concluded ironing out the performance issues was not really worth the effort.
And this brings me back to the present day, and what we should be looking for at WWDC. One thing I learned from our SwiftUI tool is that a fully declarative style presents some performance challenges. By its nature, SwiftUI is based around view tree diff-ing. When something in the model changes, it triggers a binding to fire, which in turn invalidates the view tree. In order to update the screen, SwiftUI now compares the new tree with the old, and determines what changes are needed to the scene that the user sees.
This process is fast for a few standard views, but there are bigger challenges looming. For example, take a large array of values that you want to present in a
List. We encountered some serious performance issues when updating data in such a
List, as SwiftUI would attempt to diff thousands of
View nodes. You could restructure your UI in order to avoid such large lists, but that is a bit of a cop out, because the technologies SwiftUI is designed to supplant can handle such large lists.
The problem is perhaps even more apparent in a text view. We briefly saw a
TextView type appear during the SwiftUI beta in 2019, but it didn’t make the final release. My guess — and it is just a guess — is that someone in Apple concluded that it worked fine if you were presenting tweets, but things fell apart when you wanted real-time editing on Chapter 5 of “War and Peace”.
You can bind a
TextField to a model
String property, and it will work fine, but do you really want to bind the contents of a
TextView to an
AttributedString the length of a novel? Diff-ing that is not going to cut it.
One way they could take SwiftUI in order to solve this is to introduce a new type of text model, one with internal hierarchy that makes diff-ing extremely efficient. Think about a tree-like structure, so that it could very quickly establish which subtrees were altered, and ignore all other parts of the tree. Effectively, it would be introducing a structure similar to the view tree itself, but inside the model.
This option would see
NSTextStorage replaced with a highly diff-able structure, but it may also be possible to introduce a similar approach at a different level. Perhaps the model data would be left unchanged, and a specialized
TextStorageBinding type added, which would embellish the model with metadata to make diff-ing efficient. Or could they even introduce
View types such as
Paragraph which shadow the model text content, and allow exactly the same diff-ing mechanism to be used as in the rest of SwiftUI?
Some of these proposals may seem far fetched, but this is exactly how we structured the text in the Agenda notes app. First, our model has finer granularity than just putting the complete text of a note into a single property. Each note is broken into paragraphs, meaning that a user edit doesn’t require a complete refresh of the whole note. Instead, just the affected paragraphs can be saved, refreshed on screen, and synced to other devices. This approach does require extra bookkeeping — we have to pass around the ‘edited’ paragraph identifiers — but it is this type of data that SwiftUI will no doubt need to introduce if it is ever to host an editable
TextView containing the whole of Tolstoy’s classic novel.
At this point, it is anyone’s guess which direction SwiftUI will take. Perhaps Apple figure the days of complex apps are over, and they are not going to do anything to address such concerns. Or perhaps it is just a question of priorities, and we will see how they intend to tackle these scaling issues at WWDC2020. Either way, I will be seeking out the documentation for
TextView; that’ll be my canary for where SwiftUI is headed, and whether it will soon evolve into a capable replacement for UIKit/AppKit.