Structure and Interpretation of Computer Programmers

I’ve spent about a year working on an app for a group in the University where I work, that needed to be available on both Android and iOS. I’ve got a bit of experience working with the Apple-supplied SDKs on iOS, and a teensy amount of experience working with the Google-supplied SDKs on Android. Writing two apps is obviously an option, but not one I took very seriously. The other thing I’ve reached for before in this situation is React Native, where I’ve got a little experience but quite a bit of understanding having worked with React some.

Anyway, this project was a mobile companion for a desktop app written in C# and Windows Forms, and the client was going to have to pick up development at the end of my engagement. So I decided that the best approach for the client was to learn how to do it in Xamarin.Forms, and give them a C# project they could understand at the end. I also hoped there’d be an opportunity to share some code from the desktop software in the mobile project, though this didn’t pan out in the end.

It took a while to understand the centrality of the Model-View-ViewModel idea and how to get it to work with the code I was writing, rather than bludgeoning it in to what I was trying to do. Ultimately lots of X.F works with data bindings, where you say “this thing and that thing are connected” and so your view needs a that thing so it can display this thing. If the that thing isn’t in the right shape, is derived somehow, or shouldn’t be committed to the model until some other things are done, the ViewModel sits in the middle and separates the two.

I’m used to this model in a couple of contexts, and I’ll give Objective-C examples of each because that’s how old I am. In web applications, you can use data bindings to fill in parts of an HTML document based on values from a server-side object. WebObjects works this way (Rails doesn’t, it uses code to fill in parts of etc). The difference between web app data bindings and mobile app data bindings is one of lifecycle. Your value needs to be read once when the page (or XHR) is rendered, and stored once when the user posts the changes. This is so simple that you can use straightforward accessor methods for it. It also happens at a time when loading new content is happening, so any timing constraints are likely to come from elsewhere.

You can also do it in what, because I’m that old, I’ll call rich client applications, like Xamarin.Forms mobile apps or Cocoa Bindings desktop apps. Here, anything could be happening at any time: a worker thread could be updating the model, and the user could interact with the UI, all at the same time, potentially multiple times while a UI element is live. So you can’t just wait until the Submit button is pressed to update everything, you need to track and reflect updates when they happen.

Given a dynamic language like Objective-C, you can say “bind this thing to that thing with these options” and the binding library can rewrite your accessors for this thing and that thing to update the other when changes happen, and avoid circular updates. You can’t do that in C# because apparently more typing is easier to reason about, so you end up replicating the below pattern rather a lot.

public class MyThingViewModel : INotifyPropertyChanged
{
  public event PropertyChangedEventHandler PropertyChanged;
  // ...
  private string _value;
  public string Value
  {
    get => _value;
    set
    {
      _value = value;
      PropertyChanged?.Invoke(this, new PropertyChangedEventArgs(nameof(Value)));
    }
  }
}

And when I say “rather a lot”, I mean in this one app that boilerplate appears at least 126 times, this undercounts because despite being public, the PropertyChanged event can only be invoked by instances of the declaring class so if a subclass adds any properties or any change points, you’re going to write protected helper methods to be able to invoke the event from the subclass.

Let’s pivot to investigating another question: why is Cocoa Bindings a desktop-only thing from Apple? I’ve encountered two problems in using it on Xamarin: thread confinement and performance. Thread confinement is a problem anywhere but the performance things are more sensitive on mobile, particularly on 2007-era mobile when this decision was made, and I can imagine a choice was made between “give developers the tools to identify and fix this issue” and “don’t give developers the chance to encounter this issue” back when UIKit was designed. Neither X.F nor UIKit is wrong in their particular choice, they’ve just chosen differently.

UI updates have to happen on the UI thread, probably because UIKit is Cocoa, Cocoa is appkit, and appkit ran on an OS that didn’t give you an easy way to do multiple threads per task. But this has to happen on Android too. And also performance. Anyway, theoretically any of those 126 invocations of PropertyChanged that could be bound to a view (so all of them, because separation of concerns) should be MainThread.BeginInvokeOnMainThread(() => {PropertyChanged?.Invoke(...)}); because what if the value is updated in an async method or a Task. Otherwise, something between a crash and unexpected behaviour will happen.

The performance problem is this: any change to a property can easily cause an unknown amount of code to run, quite often on the UI thread. For example, my app has a data grid (i.e. spreadsheet-like table view) with a “selection” column containing switches. There’s a “select all” button, and a report of the number of selected objects, outside the grid. Pressing “select all” selects all of the objects. Each one notifies observers that its IsSelected property has changed, which is watched by the list view model to update the selection count, and by the data grid to update the switches. So if there’s one row in the grid, selecting all causes two main-thread UI updates. If there are 500 rows, then 1000 updates need to run on the main thread in response to that one button action.

That can get slow :). Before I understood how to fix this, some UI actions would block the UI for tens of seconds as they computed the update. I asked about this in some forums and was told the answer is “your users shouldn’t have that much data in a mobile app, design an app with less data” which is not that helpful. But luckily the folks over at SyncFusion were much more empathetic, and told me the real solution is to design your views and view models such that you can turn off updates while you’re doing some big change, then turn them back on and recalculate the state at the end.

Like I say, it’s likely that someone at Apple already knew this from the Cocoa Bindings times and decided “here’s a great technology, and here’s how to turn it off because it will get in your way” wasn’t a cool story.