One question often asked about iOS (and I guess Mac, and every other UI-driven platform) is how to test UIs. A lot of us don't do it at all, often saying things like: “you should only test your business logic.” Others want to test the UI, but deem it too complex.
Whenever someone tells me UI testing is hard, I think back to something Landon Fuller said about testing the UI of Paper (by 53) during a panel about testing that we were both part of:
What you see on the screen is the culmination of a variety of data and transforms applied to that data over time ... Being able to decompose those things into testable units means you can break...down [things] that are relatively complex into more easily understood elements.
Paper’s UI is relatively complex. Testability is usually not something taken into account when building such a UI. However, any action taken by the user is modeled in code somewhere; it’s always possible to fake the user’s action in a test. The problem is that most frameworks, including UIKit, often don’t publicly expose the necessary lower-level constructs.
Knowing what to test is as important as knowing how to test. I've been referring to “UI testing” because that's the accepted term for the type of testing I'm going to discuss. In truth, I think you can split UI testing into two categories: 1) behavior and 2) aesthetics.
There is no way to deterministically say that aesthetics are correct, as they tend to change very often. You don't want to have to change your tests every time you're tweaking the UI. That's not to say that you can't test the aesthetics at all. I have no experience with it, but verifying aesthetics could be done with snapshots. Read Orta’s article to learn more about this method.
The remainder of this article will be about testing user behavior. I've provided a project on GitHub that includes some practical examples. It’s written for iOS using Objective-C, but the underlying principles can be applied to the Mac and other UI frameworks.
The number one principle I apply to testing user experience is to make it appear to your code as if the user has triggered the action. This can be tricky because, as said before, frameworks don't always expose all of the necessary lower-level APIs.
Projects like KIF, Frank, and Calabash solve this problem, but at the cost of introducing an additional layer of complexity — and we should always use the simplest possible solution that gets the job done. You want your tests to be deterministic. They need to fail or pass consistently. The worst test suites are those that fail at random. I prefer not to use these solutions because, in my experience, they introduce too much complexity at the cost of reliability and stability.
Note that I've used Specta and Expecta in the example project. Technically, this isn't the simplest possible solution — XCTest is. But there are various reasons why I prefer them, and I know from experience that they don't affect the reliability and stability of my test. As a matter of fact, I'd wager that they make my tests better (a safe bet to make, since better is ambiguous).
Regardless of your method of testing, when testing user behavior, you want to stay as close to the user as possible. You want to make it appear to your code as if the user is interacting with it. Imagine the user is looking at a view controller, and then taps a button, which presents a new view controller. You'll want your test to present the initial view controller, tap the button, and verify that the new view controller was presented.
By focusing on exercising your code as if the user had interacted with your app, you verify multiple things at once. Most importantly, you verify the expected behavior. As a side effect, you're also simultaneously testing that controls are initialized and their actions set.
For example, consider a test in which an action method is called directly. This unnecessarily couples your test to what the button should do, and not what it will do. If the target or action method for the button is changed, your test will still pass. You want to verify that the button does what you expect. Which action the button uses, and on which target, should not concern your tests.
UIKit provides the very useful sendActionsForControlEvents:
method on UIControl
, which we can use to fake user events. For example, use it to tap a button:
[_button sendActionsForControlEvent: UIControlEventTouchUpInside];
Similarly, use it to change the selection of a UISegmentedControl
:
segments.selectedSegmentIndex = 1;
[segments sendActionsForControlEvent: UIControlEventValueChanged];
Notice that it's not just sending UIControlValueChanged
. When a user interacts with the control, it will first change its selected index, then send the UIControlEventValueChanged
. This is a good example of doing some extra work to make it appear to your code as if the user is interacting with the control.
Not all controls in UIKit have a method equivalent to sendActionsForControlEvents:
, but with a bit of creativity, it's often possible to find a workaround. As said before, the most important thing is to make it appear to your code as if the user triggered the action.
For example, there is no method on UITableView
to select a cell and have it call its delegate or perform its associated segue. The sample project shows two ways of working around this.
The first method is specific to storyboards: it works by manually triggering the segue you want the table view cell to perform. Unfortunately, this does not verify that the table view cell is associated with that segue:
[_tableViewController performSegueWithIdentifier:@"TableViewPushSegue" sender:nil];
Another option that does not require storyboards is to call the tableView:didSelectRowAtIndexPath:
delegate method manually from your test code. If you're using storyboards, you can still use segues, but you have to trigger them from the delegate method manually:
[_viewController.tableView.delegate tableView:_viewController.tableView didSelectRowAtIndexPath:[NSIndexPath indexPathForRow:0 inSection:0]];
expect(_viewController.navigationController.topViewController).to.beKindOf([PresentedViewController class]);
I prefer the second option. It completely decouples the test from how the view controller is presented. It could be a custom segue, like a presentViewController:animated:completion
, or some way that Apple hasn't invented yet. Yet all the test cares about is that at the end, the topViewController
property is what it expects. The best option would be to ask the table view to select a row and perform the associated action, but that's not currently possible.
As a final example of testing controls, I want to present the special case of UIBarButtonItem
s. They don't have a sendActionsForControlEvent:
method because they're not descendents of UIControl
. Let's figure out how we can send the button action and, to our code, make it look like the user tapped it.
A UIBarButtonItem
, unlike UIControl
, can only have one target and one action associated with it. Performing the action can be as simple as:
[_viewController.barButton.target performSelector:_viewController.barButton.action
withObject:_viewController.barButton];
If you're using ARC, the compiler will complain because it can't infer the memory management semantics from an unknown selector. This solution is unacceptable to me because I treat warnings as errors.
One option is to use #pragma directive to hide the warning. Another alternative is to use the runtime directly:
#import <objc/message.h>
objc_msgSend(_viewController.barButton.target, _viewController.barButton.action, _viewController.barButton);
I prefer the runtime method because I dislike cluttering my test code with pragma directives, and also because it gives me an excuse to use the runtime.
To be honest, I'm not 100% certain these 'solutions' can't cause issues. This doesn't solve the underlying warning. Tests are usually short lived, so any memory issues that do occur are unlikely to cause problems. It's been working well for me for quite some time, but this is a case I don't fully understand, and it could turn into a bug that randomly fails at some point. I'm interested in hearing about any potential issues.
I want to end with view controllers. View controllers are likely the most important component of any iPhone application. They're the abstraction used to mediate between the view and your business logic. In order to test the behavior as best as possible, we're going to have to present the view controllers. However, presenting view controllers in test cases quickly leads me to conclude they weren't built with testing in mind.
Presenting and dismissing view controllers is the best way to make sure every test has a consistent start state. Unfortunately, doing so in rapid succession — like a test runner does — will quickly result in error messages like:
-
Warning: Attempt to dismiss from view controller <UINavigationController: 0x109518bd0> while a presentation or dismiss is in progress!
-
Warning: Attempt to present <PresentedViewController: 0x10940ba30> on <UINavigationController: 0x109518bd0> while a presentation is in progress!
-
Unbalanced calls to begin/end appearance transitions for <UINavigationController: 0x109518bd0>
-
nested push animation can result in corrupted navigation bar
A test suite should be as fast as possible. Waiting for each presentation to finish is not an option. It turns out, the checks raising these warnings are on a per-window basis. Presenting each view controller in its own window gives you a consistent start state for your test, while also keeping it fast. By presenting each in its own window, you never have to wait for a presentation or dismissal to finish.
There are more issues with view controllers. For example, pushing to a navigation controller happens on the next run loop, while presenting a view controller modally doesn't. If you're interested in trying out this way of testing, I recommend you take a look at my view controller test helper, which solves these problems for you.
When testing behavior, often you need to ensure that, through some interaction, a new view controller was presented. In other words, you need to verify the current state of the view controller hierarchy. UIKit does a great job providing the methods needed to verify this. For example, this is how you would make sure that a view controller was modally presented:
expect(_viewController.presentedViewController).to.beKindOf([PresentedViewController class]);
Or pushed to a navigation controller:
expect(_viewController.navigationController.topViewController).to.beKindOf([PresentedViewController class]);
Testing the UI isn't hard. Just be aware of what you're testing. You want to test user behavior, not application aesthetics. With creativity and persistence, most of the framework shortcomings can be worked around without sacrificing the stability and maintainability of your test suite. Just always remember to write tests to exercise the code as if the user is performing the action.