Hacker News Reader

Comment by jasonkester | Hacker News Reader

Here are 4 real-world bug reports, in their entirety. Anybody care to rewrite them as automated test cases for me?

  * "Try it Now" button sometimes drops below the "Learn More" button in FireFox

  * Capitalization inconsistent in top nav

  * Popup blockers occasionally stop the App window from opening properly

  * Wierd timeout errors clumped around 1:00 am every few nights

  * Accounting needs to be able to call up weekly income reports

In my experience, it's actually pretty rare to find a bug that can be wrapped in a test case without actually discovering and fixing it in the process. Most things that end up in the bug tracker are either cosmetic, wierd "solar activity" things, or disguised feature requests.

View on HN

Whenever the unit testing comes up, people always cite GUI issues as untestable. But it should be pointed out that in Fred Brook's usage of the terms, this is an accidental problem, not an essential one. GUIs are hard to test not because GUIs are hard to test, but because your GUI doesn't contain provisions for being tested.

Nothing stops a GUI from being very queryable. Nothing stops you from being able to query the locations of two widgets and making assertions based on those locations. Nothing stops you from getting text back out, or verifying fonts, or verifying the lack of overlap, or lack of popups... except that GUIs are written to be opaque monoliths, graveyards of data. This is correctable, but not by the end programmer.

So it is true that GUIs are untestable, but this doesn't have to be true. But it's going to take a conscious effort by QT or GTK or someone to make their GUIs testable before anything will change.

(I acknowledge your other issues; don't agree with all of them but the GUI issue is what I'm passionate about.)

> So it is true that GUIs are untestable, but this doesn't have to be true.

I respectfully disagree. You can't codify aesthetics or usability, and even if you could, the return on investment would be abysmal.

"Aesthetics" or "usability" as a whole? Sure, of course not. No more than you can codify "works correctly" as a whole.

Codify huge swathes of it, though? Absolutely. Assert that correct fonts and font sizes are used. Assert that color schemes are correctly followed, or that certain widgets always have attached desirable behaviors, assert that all widgets have keyboard shortcuts, assert that those keyboard shortcuts are coherent (CTRL-F is not assigned to three actions), assert that using the keyboard shortcut has the same behavior as whatever it is the shortcut to, assert that automated validation of appropriate accessibility guidelines hold (ie, for web interfaces assert that a Bobby run over your interface holds no surprises), and so on.

Sure, you can't "automate usability" but if you just give up at that point you'll miss out on a lot of things you can automate... or could automate if GUIs let you, getting back to my original point.

All of the examples you give are functionality, not how the thing actually looks when it renders on the screen.

Is there some way to write a test that determines whether my page renders correctly on Firefox, Chrome, Safari, IE 6, IE 7, IE 8, and Opera? If so, please show me the way.

In web development, these sorts of things end up being a large percentage of bugs filed in FogBugz, and there's no way to test for them but to have someone look at the page and make sure.

We had manual test case for "page looks correct" test. It consists of more than 10 checks, i.e. "no JavaScript errors", "no miss-spelled words", "consistent look across pages", "no broken links", etc.

More than half of that list is already automated.

You are right of course, but nobody said that GUIs weren't testable. Just that some bugs cannot be expressed as tests.

To do be basic GUI testing, you don't actually need to test aesthetics, just some basic morphological/topological perperties like "is this button touching that button", "is this window visible?" etc.

But this is not likely to happen either.

Part of the problem with making a windowing system query-able is that your talking about a number of distinct components that each separately paint themselves onto the screen. The screen's properties vary by system's configuration, the graphics driver, and the size of the overall window being painted. Sure, all of those components could be rewritten so at least relationships like "above" and "beside" and "touching" could be queried. But this stack involves hardware vendors, OS creators and software tool vendors. Moreover, hardware vendors and OS creators wouldn't want add anything that got in the way of performance since that's their main claim to fame.

As a Rails developer doing a fair amount of TDD and some BDD, I have a first-hand view of the state of the art of view testing for the web. Tools like webrat, selenium and cucumber are compelling because they hold the promise of end to end testing.

I'm really glad that people are working on solving these hard problems, but, I'm also a designer and usability guy, and I can tell you that these tools are still very very crude. You know the sayings "when all you have is a hammer everything looks like a nail", and "use the right tool for the right job"? Well, to developers, code is the hammer. Developers (especially consultants) would like to believe that if they write a cuke test and it passes, then their job is done. But the fact of the matter is that test code is not free to write or even to run, and the amount of things that a QA person is actually testing when they click through an application is several orders of magnitude more coverage than an state-of-the-art acceptance test gives you (this may change eventually, but it's an AI-level problem).

I see a lot of value for the initial acceptance tests that verify basic flows, but they are just one tool in your belt. Usability testing and QA provide orthogonal human perspectives that provide more bang for your buck then increasingly fine-grained automated tests. Think of it like this, automated tests take your product from complete failure to mediocre, but design, usability testing and QA can take it from mediocre to great.

I've seen brilliant programmers get obsessed with "all green" to the point that they do all their work without firing up a browser, and miss dozens of terrible usability flaws that would be immediately obvious the first time you click through.

None of this is in any way disparaging of GUI testing per se, but just that it has limitations which I've seen ignored for the sake of code fetishism.

Not sure if it's any different or better, but Project Sikuli (http://groups.csail.mit.edu/uid/sikuli/), an api for scripting visual actions in guis could be useful for testing layout and interaction.

Here's an example of using it for unit testing a gui: http://sikuli.org/documentation.shtml#examples/TestJEdit.sik...

The way I see it is that the code to test if a GUI element is properly formed is basically the same code that would just properly form the element in the first place. So by automatically testing the GUI you are basically just repeating yourself over and over.

Not to mention that when you want to improve the usability by rearranging things, it will be twice as hard to do if you need to change both the code and the tests.

Unless you have to render something pixel-perfect, it is probably not worth it to write automated tests. A human can look at your screens and verify that they aren't (or are) completely crap fairly quickly.

Ultimately humans will be the ones finding bugs in your software, and when they find UI layout bugs it should be immediately obvious by looking at the code what is wrong. If it isn't then you might need to separate UI from the rest of the gunk better.

Yes, bugs still need to be reported, and not all of them yield to automated tests. Nevertheless I do concur with the spirit of the article. I write very demanding test cases, often before anything goes wrong but certainly after it does, even for non-deterministic code (e.g. I'll fork a thousand processes in an automated test and invoke random busy loops here and there to encourage contention.)

I don't do a lot of GUI though, other than HTML interfaces, which are fairly amenable to some automated testing. One time I broke a sign-up form due to a misnamed field, so now I have an automated test for that.

For things like "Capitalization inconsistent in top nav", that's a prime example of why you still need to keep track of bugs. Even if you're working alone, just keep a "TODO" list. You can keep the TODOs in a note-taking program such as Tomboy, but better yet, put them right in the code. Then don't release until "grep -r TODO" comes up clean. (If you want to postpone a TODO change it to LATER.)

It might be difficult to create a regression test for that one, and it probably wouldn't be cost effective, but I suppose could have a snippet of code that iterated over the list of navigation items and tested the case. Again, probably not worth it. Better just to stomp that TODO and move on.

I can't say I have the silver bullet for the examples you list, but the important thing is to keep manufacturing better bullets with higher silver content. I figure I gotta try at least.

different programming domains have different problems...

wrt this list:

  * if GUIs were easier to write tests for, then you wouldn't have problems writing tests for the first 3 on your list.

  * timing issues and edge cases are one of the reasons i value tests, assertions and strong design so much: debugging is hard.

  * the last one sounds like a feature that hasn't been written, not a bug.

timing issues and edge cases are one of the reasons i value tests, assertions and strong design so much: debugging is hard.

It always seems like when you scratch the TDD surface, "writing tests first lets you write bug-free software" becomes "tests are an important tool along with good design, assertions, etc". I agree with the later but I don't see how it's different earlier methodologies - and earlier methodologies certainly didn't let you write bug-free code.

Check out Selenium.

> * "Try it Now" button sometimes drops below the "Learn More" button in FireFox

Use JavaScript or Selenium for that Test case. Get coordinates of buttons (getBoundingClientRect()) and compare them. Execute this test case on tons of different configurations.

> * Capitalization inconsistent in top nav

Easy to test. Also add spell checking along the way.

> Popup blockers occasionally stop the App window from opening properly

Selenium?

> Wierd timeout errors clumped around 1:00 am every few nights

You always can shift clock automatically (on Linux).

> Accounting needs to be able to call up weekly income reports

If it is missing feature, then no test case is need until you will start work on it.

PS. If you want to automate your manual testing, you will look for ways to do that. If not, you will look for problems, which are prevents you from doing that.

I wrote example how to execute HTMLUnit tests at build stage because I want that. Almost everybody else says that this is not possible. ( http://vlisivka.pp.ua/en/integration_testing_with_maven_test... )

> Wierd timeout errors clumped around 1:00 am every few nights

You always can shift clock automatically (on Linux).

Obviously a problem of this sort is not due to the mere numerical value of the time but rather some daily scheduled process that runs around 1am (perhaps a database backup) which interferes with the web app in some way. This sort of thing can be simulated but it's a non-trivial task and it's probably better to apply those efforts in other ways (by diagnosing the problem, putting in place a performance monitoring system, buying more hardware, etc.)

Typical QA cycle:

  * find problem;
  * identify root cause of the problem;
  * write automated, semi-automated or manual test case, which will highlight problem;
  * fix problem.

If you will not use test cases at development and staging stages, then you will just use your production environment for testing purposes. If your downtime cost you nothing - don't bother with test cases.

In our case, we had similar problem, so I created very simple test case in shell in less than 2 hours. We spent more than year until it was fully fixed.

Reply on news.ycombinator.com