100,000 e2e selenium tests? Sounds like a nightmare
watirmelon.com
watirmelon.com
Huh, after working on very large applications with tens of thousands of unit tests I'd choose the opposite.
I've found little value from unit tests. They're sometimes good enough for catching the lowest common errors. With very complex applications that involve a great deal of user input they wind up only being a CRUD layer test.
I've gone the salesforce route, instead of using selenium all my "unit tests" are executed against the rest api.
In order to test "Deleting a product from the system" I have to create a user, assign the user as an administrator, log off of the root admin, log in as the new user, create a product, delete a product, get the product list and verify the product isn't on the list.
Typically this would be answered in a unit test like this:
p = productDao.create("My New Product")
assert true p.delete()
ymmv
Integration test the happy paths, small integration suite that proves everything works together.
Unit test the individual units of functionality or behavior, whichever you're more comfortable with. Unit tests aren't supposed to tell you that the whole system works. Unit tests exist for two main reasons: Confidence that you can refactor your code without breaking functionality, and proving that your components are doing what they're supposed to do. Yes you technically can accomplish this with just an integration test suite but you will find refactoring bugs much harder to track down because the test failures aren't telling you exactly what went wrong.
You can never have full confidence in a system with just one suite of tests. You really need multiple suites. And when you TDD it all the way down (Integration test first, then work through the stack TDD-ing each part, when done Integration test and all unit tests should pass, and feature is done!) it's not a chore but actually quite enjoyable.
Also, Unit tests are where you ensure the non-happy-paths are functional (error handling, input robustness, etc).
Edit: "Typically, refactoring applies a series of standardised basic micro-refactorings, each of which is (usually) a tiny change in a computer program's source code that either preserves the behaviour of the software, or at least does not modify its conformance to functional requirements." http://en.wikipedia.org/wiki/Code_refactoring
Refactoring without introducing bugs is the #1 reason people do TDD vs Test After. But if you don't trust the test suite to catch bugs introduced when refactoring then I'd argue the test suite itself probably isn't that useful and needs work.
On a side note I believe the term "refactoring" has been way overused these days. You can change code without it being "refactoring". When I've seen people use the term "refactoring" (and I've caught myself multiple times on this) they mean rewriting, for in most cases this rewriting means changing the code and changing the tests. If you have to change the tests because of changes to the code, that's not refactoring, that's just changing code.
Refactoring is the process of changing a software system in such a way that it does not alter the external behavior of the code yet improves its internal structure.
You correctly identified a consequence of this in your later comment:
> In your definition if you introduce or remove a bug your not refactoring.
That is indeed the case.
And also correctly observed that:
> realistically for any sufficiently large change bugs can and will be both added and removed.
If a change adds or removes a bug, it was not a pure refactoring. It may have been an attempt at a pure refactoring, and it may have been otherwise successful, but the introduction of a change in behaviour means that it was not purely a refactoring. This isn't necessarily a bad thing - i'd rather bugs be removed than not! - but it's possible to distinguish the refactoring and bug-fixing aspects of the change. Ideally, i would like to see those aspects formally separated, for example into separate commits in source control. But even if not, we can at least use the terminology correctly and precisely.
Granted, that may seem pedantic but if you look closely you realise nobody uses your 'ideal' definition in practice.
PS: Feel free to use / introduce new terms such as 'Pure refactoring' but understand they don't change what refactoring actually means. As to popularity wikipedia's "does not modify its conformance to functional requirements" suggests that's the commonly understood definition.
And Unit tests alone cannot ensure that the non-happy-paths are functional.
Interesting. I think this is probably right and also the reverse of what most people try to do (they start with units and work up)
- it forces you to flesh out your conceptual/paper design in a deliverables-oriented way and makes it obvious where the gaps are, which gives you a good idea of how ready you are to wrap up design and start development, and in turn can be used to create your work breakdown structure and task prioritisation; sometimes it can result in identifying simplifications upfront (saving you time down the line)
- when you have a good idea of the scope and complexity of a solution like this, you're less likely to waste time on the less important stuff
- if you're the tech lead on a team, giving your developers the high-level components and tests means they're more likely to get things right (or at least limit the damage they do when they get things wrong)
Without component-level unit tests or design, you can easily dive into a solution without thinking it through and really regret it.
Integration testing should, at a minimum, include:
* Combinatorial testing
* i.e. testing all possible input combinations for every pair of two inputs
* Fuzz testing
* Limit testing
* Success path testing
* Failure case testing
On the other hand, I wholly believe that testing done by the developer alone will never catch as many bugs as testing done by both the developer and a QA engineer to the mix will. They think differently than we do, and as such are adept at finding our blind spots.
Agreed having a QA team also go through the app is great for finding things that tests can't catch, like usability errors, or crazy edge cases you didn't think about, etc. I have gotten to work with a good QA team and it's amazing the things they find.
Whenever you fix a bug in your code, try to cover it with a unit test. If you can't, ask yourself whether it might be because the code is poorly abstracted.
I think there are definitely varying degrees of value in unit tests. Your example is definitely not super valuable, but unit tests on higher level methods can save a lot of trouble. When I code I want to rest assured that the different pieces I am using work. I want assurance that this class I'm about to use has an API that does exactly what it says. If I want to change modify the class, I want to see exactly what assumptions I'm breaking that users of this class might be making, and I want to be able to do this in under a second. This can have a drastic impact on developer velocity. I don't know of any other way of doing this than unit tests.
This, big time. Whenever I get a bug report I first write a test that fails because of the bug. Only then can I a) be sure I understand the bug and why it happened and b) be sure I fixed the bug and finally c) prevent a regression in the future.
Even for people who are not TDD or big on tests in general can make great use out of writing tests to verify and fix bugs.
Not even when re-factoring? Refactoring code without unit tests makes me extremely anxious and I try to avoid it.
I've done major refactoring (moving from one language to another) and the testing was done entirely on the rest api. With several thousand tests I felt very confident that the user experience and expected behavior remain unchanged.
Honestly I don't hate unit testing and do I use it on a regular basis, the article was rather heavy handed.
(caveat: i understand "unit test" to mean a test which tests a single class, or sometimes a very small number of classes, with any collaborators replaced with stubs or mocks)
In my experience, bigger tests - what my colleagues call integration tests, which might involve 3-30 classes and often the database - give me a lot more confidence in my changes.
Tests are only a useful safety net for changes which are entirely confined to the thing they test. If you're changing the way some particular method is implemented, say from looping over a collection to mapping a lambda over it, then the change is confined to that class, and a unit test can help. But if you're changing the method's contract, say from taking a collection of objects to taking one at a time, then the change takes in that class and its clients, and the unit test is useless - it tests that the method does something that it should no longer do! Yes, you can rewrite the unit test, but that does nothing to reassure you that the overall behaviour hasn't changed. You're starting from scratch.
The thing is, in my experience, valuable refactorings tend to take in more than one class. Sometimes much more than one class. I want to be able to make those refactorings with a safety net. Unit tests can't give me that.
Where testing is useful is when there's complex logic. And such logic is much easier to test at the unit level.
p = productDao.create("My New Product")
assert true p.delete()You, on the other hand, argued that simple deletion didn't need ANY tests, unit or functional - I was responding with examples that demonstrated that functional integration tests for deletion are perfectly valid.
This would be a really bad way to test all of "create a user, assign the user as an administrator, log off of the root admin, log in as the new user, create a product, delete a product, get the product list and verify the product isn't on the list".
You would be ignoring 80% of the flow.
The goal of unit testing is to test each piece in isolation and to provide confidence that when you have to assemble a lot of little pieces of code to form a larger piece of functionality, that each of those little pieces is working as it's contract says it should.
If your integration test containing 8 steps failed, would you know which piece of the puzzle caused the failure from the integration test report itself?
Of course, you could go the route of testing both small pieces of functionality and also testing large pieces (the integration tests), so that you can have overlapping levels of confidence.
In my experience, the value in unit tests is in refactoring. I'd only write them as I refactor [to confirm the functionality matches before/after the refactor].
Most of the modifications we do to common libraries, etc. involve higher level rewrites that would break the unit tests anyway and the only way to confirm things still function is integration tests. :/
I've taken the same route for apis and for client side testing.
On my current project, I'm creating a rich client. I have an extensive jasmine test suite with edge case and race condition testing.
I created a library called jasmine-flow, to organize the tests into flows. This allows edge case testing while reducing duplicate test setup. In my case test suite time was reduced by 10x.
I agree with the overall thrust of the article. Unit testing is indispensable, especially for me as a rubyist where there is basically nothing checked before runtime. But I find this conclusion that SalesForce is "doing it wrong" to be just as arrogant and presumptuous as DHH's blustery approach.
Now I don't know anything about SalesForce's architecture, but one thing I do know is that no system of that size exhibits the same ideals that can be maintained by a small team of brilliant developers working on the bleeding edge of some greenfield project with a set of stakeholders that fits in a single room. That's not to say that SalesForce couldn't use unit tests, but just that it's the height of arrogance for us as outsiders to presume one way or the other.
Scaling software development isn't solved by strict adherence to the dogma of idealistic methodologies any more than scaling traffic is solved by plugging in a "web-scale" nosql database into your Heroku. The reality is that scaling is done by removing bottlenecks over and over until the system that remains is an organic result reached through evolution rather than any architect's brilliant grand design.
Considering this is a business that is spending vast sums of money on doing this I think you are spot-on, yes we all joke about businesses throwing good money after bad but someone had to sit down and defend this approach.
1) Use the page object model. http://martinfowler.com/bliki/PageObject.html
2) Think about your functional tests like tours through a city. http://ptgmedia.pearsoncmg.com/images/9780321803023/samplepa...
3) As your tests become more advanced abstract them locally to two simple questions. What page object did I come from, and what action did I do on that page to get to this page?
The real problem with end to end tests is that when end to end tests fail, most of the time you have no idea what went wrong so you spend a lot of time trying to find out why.
I've worked on big systems with predominantly (in some cases, exclusively) high-level test coverage. I can confirm that this indeed a thing. Unit tests very rarely fail in such a way that the cause of the failure is not immediately obvious. High-level tests often do. Not most of the time, but often.
Here's the thing, though. When a high-level test fails without an obvious cause, that is something that unit tests would not have caught at all.
Obvious causes are things that are localised defects in individual parts. Your product listing is showing in the wrong order? The error is in the sorting code. Orders in Japan are being billed with prices off by a factor of about a hundred? The error is in the currency conversion code. Unit tests would catch those errors and make their location blindingly obvious. High-level tests catch them, and make their location sufficiently obvious; obvious enough that pinpointing the problem is not a significant speed bump.
Non-obvious causes are things that are defects in interactions between multiple, often distant, parts. Only every second FTP'd file is triggering emails to users? Turns out the error handling code in the parser is disabling the scheduled refresh checking job, which is then not triggering the batch job to recalculate the numbers, which is then not feeding new data to the email job. Every single part is working as intended, but the composition of those parts is not. That's a tough problem, and when the high-level test catches it, you will have hours and hours of the wrong kind of fun unravelling the strands of cause and effect to pinpoint the problem. But unit tests would not have caught it at all.
On my current project no one would dream of adding a new feature without unit tests - but everyone turns a blind eye to integration and system tests because "they're hard".
The higher you get the more time it takes for tests to run, the more difficult it is to identify where a failure is, the more brittle those tests will be because of inter-dependencies, and the more complex it is to write the test. The difficult part is understanding your system well enough to know where your unit tests aren't sufficient.
The slide "Why Salesforce loves WebDriver" explains it perfectly. It's a feature--not a bug--that they experience latency (users do too and latency is realllly important for web apps).
It is a lot of VMs, but Salesforce is a $32B company entirely based in the cloud, a few million bucks of metal to test their breadwinner seems cheap.
I've been doing this for years. It's been working out very well. I'm glad test abstraction scope is being talked about now.
Test scope has been talked about for years, but almost always in the form of sermons by fervent believers in unit testing, castigating those who dare to lean more heavily on system and integration tests for their dangerous heresy.
Occasionally, these sermons admit the existence of a "testing pyramid", but take note of the necessary shape of that edifice: it's unit tests, but with a decorative cap of other kinds of test inexorably dwindling in volume as they reach higher levels. I'm not interested in this all this alleged pyramid science. Talk to me about testing towers, testing dolmens, then we can have a genuine conversation.
Tests aren't there to test our code, but to make our Users happy. e2e tests make sure we don't break features and are able to validate new features don't kill the most common workflows. Thus we can innovate quickly, build new features and refactor. This makes our users happy. Unit tests, controller tests or others are mainly an optimisation as writing an e2e test for every failure state is not an option.
It's a different understanding of why we test, but makes sure we focus on the right things with it.
To my eye, all those unit tests are expect(true).toBeTrue() and no more!
The benefits between unit tests and e2e system tests are varied and surely that's the main benefit?
In my personal experience in software systems is that the bottom is not stable you never end up selecting a material and having component that never changes after you build it. All layers of an application tend to change as the application changes and user needs change. Therefore you can't really say that oh we tested the bottom of the pyramid and know for sure that it works.
Writing end to end tests for application is quite hard work and requires a lot of though to design an application that can tested both at the unit level, the component level, and the system level.
Even though end to end testing is very hard it's value is massive as an industry we should be focused on lowering the cost of end to end testing rather than saying that unit is good enough.