1. As your integration tests will both more numerous and complex to cover corner case scenarios, the time it takes to execute the suite will grow too. And since the integration test is more opaque too, this price has to be paid by both you and all the developers that touch something that might have used your module.
2. The time at which you receive the feedback of the test is already too late. You are no longer doing experimental changes, you are trying to push this unit of work into the code base and be done with it. Either you will be tempted to gloss over any divergence you find (quality price) or you will have to execute some relevant subset of your integration testsuite during your coding/OODA loop (execution speed price).
Choosing unit tests because of speed is nearly always a premature optimization.
IME everywhere that I worked with that relied exclusively on unit tests for automated regression testing didn't trust them enough to do releases without some manual user testing, so the project fell back into waterfall mode.
Integration tests tested a lot more realistically and gave more freedom to refactor technical debt (since changing the code didn't require changing the test as often).
Unit tests cannot catch everything though, and for stuff like how multiple units are wired together, that's where integration tests come into play.
A good codebase has both types - even combined they won't catch everything, but they should capture almost all business logic issues.
However, a lot (perhaps most) code is more about integrating disparate components (browsers, lots of different kinds of external APIs, databases, times, distributed task processors, etc.) rather than lots of complex logic and for that kind of code, unit tests are substantially less effective and more horrible to use whereas integration tests really shine.
I think ideally, you should keep the two kinds of code strongly segregated, with clean APIs between them, so you can easily use the appropriate kind of testing for either one.
I don't agree that a good code base necessarily has both types. A web app with relatively shallow logic and lots of integrations only really benefits from having integration tests. A library to convert human readable strings into dates likewise gains nothing from having integration tests.
Similarly, a messy spaghetti codebase with a disgusting mix of both types of code can only really benefit from integration testing until it's substantially refactored and the two types of code are segregated.
Why _strongly_ segregated?
Nothing stops you to write unit tests for the methods that should be covered by unit tests and integration tests for the functionality that needs to be covered by integration tests.
How would code segregation help?
That test will also often break as you refactor the various connecting APIs it surrounds.
Tests that break in the presence of changed API contracts (as opposed to bugs) are vastly less valuable than tests which only go red in the presence of a bug.
I normally structure two type of integrations test: 5-10 minutes tests that run before every checking.
Over night, over weekend tests that runs every night and every weekend.
5-10 minutes are enough time for 200 UX features level tests for both embedded system and Web backend/frontend integration type project, for me,
BTW, ALL checkin code should be passing ALL integration tests. If they don't, the highest priority is fixed the test failures and not develop any new feature.
An analogy is that unit tests are like a brick layer's spirit-level and plumb-line. As you build the wall, you ensure the pieces are put together correctly and head off mistakes early.
The "only test at a high-level" is like saying don't confirm the bricks you are laying are correct, instead bring in an inspector to check the whole building is up to code once we have finished.
If we do bring in the inspector she takes one look and says the wall have to come down because of faults right at the bottom or because we forgot the damp course or we never get that far because we couldn't fit the roof because all the walls bowed out.
I fear that those who advance anti-unit-test views only do so because they lack experience with large complex code-bases and that they are not in a position to be advising others.
And this is basically the only time I would write an unit test - when I can not capture most invariants in the types. Everything else is basically double the work and a waste of time.
Integration tests are always a good idea, but hard to get "conceptually" right. When you get to the level of testing multiple components through the whole stack most of the time IO is involved and this gets hairy fast, not to mention that one mostly tests the 3 codepaths they can come up with.
Maybe a DSL/swappable interpreter approach is one way do that? Or make everything pure again with FRP?
I can't see how exhaustively testing a small unit like this with high-level tests would be practical in terms of setup effort and run-time.
The amount of edge cases Quickcheck catches is beyond the most comprehensove unit test suite.
On a higher level though I would use parser combinators. If you have a grammar specification you can then directly transform it into a parser and thus would only need to test that you transformed it correctly (i.e. prove the proof as opposed to prove the implementation if that makes sense).
Basically - try to capture most of your invariants in the types and then you know your code is correct. The invariants may be wrong though, in which case integration testing may help.
However, I still have no idea how to test programs with complex state and IO in a way that isn't basically "script the 5 behaviors I can come up with and pray for the best", in which case I mostly don't even bother (i.e. UI etc...)
(I know this is a very short answer but I'm typing on a phone)
That's funny. I suspect that people who advance the pro-unit test view do so haven't had sufficient experience with large complex code bases.
Experience with large complex code bases with a spaghetti problem is the exact thing that put me off unit tests.
Unit tests are by their nature very tightly coupled to the underlying implementation, which is why they cause so many problems.
It's kind of like pouring concrete over a bunch of code bricks: if those bricks are in the wrong place and need refactoring, by pouring unit tests over them means you're solidifying the broken architecture and making it even harder to fix.
IME there basically don't exist large projects without technical debt problems either, so arguing that unit tests work well when the architecture is good is kind of a moot point.
> Unit tests are by their nature very tightly coupled to the underlying implementation
No, not "by their nature". Unit tests are supposed to test that a "unit" has the intended external behavior. It shouldn't care about the internal behavior of the "unit". That's the theory. They're only supposed to be coupled to the interface, not to the implementation.
I've found this to be one of the biggest myths of unit test driven development. Developers who have a tendency to write spaghetti do so anyhow. When it becomes too hard to write tests they simply don't write tests or they write 'pretend' tests.
It's perfectly possible to write spaghetti code and write a badly written test that notionally covers it.
Good coders code well. Bad coders don't. TDD doesn't help here and if there is any relationship between TDD and good design, I'm certain it's correlation not causation.
>That doesn't help you if you are adding tests to an existing project that's a mess, though.
Harms you deeply, in fact.
>No, not "by their nature". Unit tests are supposed to test that a "unit" has the intended external behavior.
i.e. it couples itself to a unit, which is an underlying implementation detail.
If it's coupling itself to an externally facing interface it's an integration test or an end to end test.
>It shouldn't care about the internal behavior of the "unit".
No, but then again, since "units" are supposed to be the smallest blocks of code that comprise your code base, it's caring very much about how all of the blocks talk to one another, which is an implementation detail (for your project). Perhaps the most crucial implementation detail, in fact.
How those units talk to one another (their API contracts) is basically the most critical aspect of your project's architecture and the single biggest source of unavoidable technical debt.
Pouring test cement in those cracks means that any technical debt accrued in those API contracts becomes impossible to refactor without breaking the tests.
In practice this makes that refactoring more difficult than if it had no tests at all due to the massive psychological inhibition against deleting tests or turning the build red and the added problem that real failures look the same as changed API contracts.
Absolutely.
> Good coders code well. Bad coders don't.
Also true.
> TDD doesn't help here and if there is any relationship between TDD and good design, I'm certain it's correlation not causation.
I've used TDD and not used it. And when programming using TDD, my code was less "spaghetti" - more modular, better-designed interfaces. There definitely was causation, in my case - it made me write my non-test code differently. So that's anecdote and not data, but holding the programmer constant and changing the methodology, the style and quality of the code changed.
> >No, not "by their nature". Unit tests are supposed to test that a "unit" has the intended external behavior.
> i.e. it couples itself to a unit, which is an underlying implementation detail.
> If it's coupling itself to an externally facing interface it's an integration test or an end to end test.
Let's get specific, so that our terms can be better defined, and we can make sure that we're not just arguing about definitions of terms. Instead of "unit", let's say "C++ class". Now: By my definitions, my unit test should test the behavior of the public API of that class, not the private implementation. So the API is an implementation detail of the project, but not of the class. I will agree that the unit test definitely couples to the interface of the class. Done right, though, it doesn't couple to the implementation of the class.
> How those units talk to one another (their API contracts) is basically the most critical aspect of your project's architecture and the single biggest source of unavoidable technical debt.
A bigger source of technical debt than the bugs you can't find that are buried in the implementation of some class? I doubt it. And good tests mean that those bugs don't make it very far before getting exposed.
A bigger source of technical debt than that API that you don't know how to use, and so you call it wrong? Tests can actually serve as documentation for how to use an API. You have non-test documentation? You're lucky to have that, but the problem is, when the code implementing the class changes, the documentation might well not. If that happens with the tests, you find out the first time you run them.
In almost all cases (the exception being if the class is the core of a library's externally facing API), the class itself is an implementation detail.
>A bigger source of technical debt than the bugs you can't find that are buried in the implementation of some class?
Yes, obviously. Bugs are not technical debt.
>And good tests mean that those bugs don't make it very far before getting exposed.
And integration tests are generally better at finding said bugs than unit tests since they test more realistically.
>A bigger source of technical debt than that API that you don't know how to use, and so you call it wrong?
Is again, a bug - not technical debt.
Unit tests are horrible for this since they require you to mock APIs that you're calling. The test which mocks the API wrongly won't catch that bug but the integration test that uses the real API will.
>Tests can actually serve as documentation for how to use an API.
Yes they can. Integration tests can double as a specification for an entire application which is even better than a documentation for a (potentially) badly architected API.
What? On the codebase I work on, integration tests run every single commit, just like unit tests. They would both fail at the same time. Except the integration tests are much less brittle.
The faster your test suite, the more often you can run it, and the faster you get feedback on your code.
You can combine this with a healthy integration test suite, and only run the unit tests.
Validation rules is a good example, if I want to test that my user input validation rules are correct, I don't need to do this through the entire application frontend - I can just check those rules in isolation.
In all likelihood, the faster it is, the more tightly coupled it will be. You're simply exchanging computing power for brain power (refactoring tests when you refactor code), which is nearly always a bad trade off.
In most cases I want to preserve existing behaviour, so I don't want any existing tests to fail.
I'll add tests for the new behaviour in the relevant modules, add code to make them pass, and then if necessary update the glue code / integration tests which cross the boundary.
Pretty much by definition, a refactor of code should not change its behaviour. My tests usually assert on behaviour observed at the boundaries of modules, so they should still pass during refactorings.
Quite precisely by definition, perhaps for other reasons as well :-P
> My tests usually assert on behaviour observed at the boundaries of modules, so they should still pass during refactorings.
I find the kind of refactoring that most desperately needs the support of tests is changing the boundaries of the individual units. Unit tests can still arguably be useful there in making me clarify the changes I'm making to my interfaces... but they do need changing, and I can't be quite as confident that nothing is breaking if I'm changing both code and tests in tandem.
I'm with GP about integration tests.
Sometimes I'll wait until I've finished changing all those function signatures until I hit save.
No it's not. Not at all. The whole point is you can change the guts of a unit around and know you haven't broken things cause the unit tests still pass. Because the unit test is testing the "contract" (explicit or implied) the unit has made with other units about it's inputs and outputs.
Unit test should fail (and require updating) only if the unit's INTERFACE changes.
But, if your unit tests are failing everytime you touch your code then 1) your code is not cleanly isolated and independent and/or 2) you have poor unittests.
Refactoring is defined as the process by which you restructure code without changing its external behaviour. So in other words, refactoring should never break tests because the external features of the code do not change.
Now - redesign or rewriting code is a very different thing. Refactoring can lay the groundwork for redesign or component level rewrites. But the process of redesign and rewrites could result in breakages unless the right level of isolating layers are put in place. This usually involves lots of work to keep tests working.
Personally I have found that unit tests don't result in less work. They are a LOT of work and will continue to be a lot of work if they are to generate value. What they offer though is the ability to know when things have gone wrong and the confidence to release and deploy. I realize that you don't get increased confidence in your release process by doing less work. Software it appears, like almost everything else, obeys the basic laws of thermodynamics. :)
You only need that sort of test in a programming language or public library, because you're publishing the behavior to unknown numbers of users, and it has to work, as documented. If they find too many things that don't work according to the documentation, they will correctly suspect that you aren't testing, and lose confidence.
But if some API is internal to an application, and no code depends on some aspect of its interface contract, then that aspect effectively doesn't exist.
It may happen that someone discovers this aspect in the documentation and tries using it. At that time it is either found broken, or found working. If it is found broken, it can be fixed. Or else it can be striken from the documentation (and another way can be found in the application code to meet its requirements).
If the aspect is found correct, or fixed to reflect the documentation, and is then used in the application, then that use constitutes a test. The feature is tested indirectly, through some regression test case which targets something in the application, which depends on that aspect of the API contract lower down.
If you design units with all kinds of detailed aspects in their interface contracts which end up not being used at all, and test all of them, it can be argued that you're wasting time: you're basically working on your library design hobby, at the expense of the project, which just serves as a vehicle for it.
"Can't tell unless" means exactly the same thing as "can tell". You have a way to get to that corner case through the public functionality.
If there really is no way to tell through the current public API, there the bug doesn't exist.
Are you arguing that all unit tests should be black-box?
And what are those "things" that haven't broken? Why, unit tests.
As mentioned, the article is directed primarily at end-to-end tests. It concludes speaking favorably of some integration tests, but says that as a guess, you want 70% unit tests, 20% integration tests, and 10% end-to-end tests. It does not actually directly address why you mightn't want more integration tests than this. I'm not at all confident that that's wrong, but I suspect that the precise ratios are context specific, and in any case it's not well supported.
The strongest point the article makes against end-to-end tests is that it can be difficult to pin down the exact cause of the error. In my experience this is certainly the case, and is less true of integration tests, and still less true of unit tests, when those tests catch the defect in question. It is less true still of types, when the type checker is sufficient and the software is designed to leverage it well.
To the weaknesses of the article with regard to its critique of end-to-end tests:
> Let's assume the team already has some fantastic test infrastructure in place. Every night:
Hold on! What we're examining now is the difference between low frequency feedback and high frequency feedback. Depending on context, we may be able to run end-to-end tests every commit. And again depending on context, unit tests may be slow to run - we can always throw more generative tests at any SUT, and exhaustive tests may give us a high degree of confidence in exchange for a large amount of compute. The chief difference here is not the scope covered by the tests.
> To maintain a high bar for product quality, they also require that at least 90% of their end-to-end tests pass before features are considered complete.
It may be desirable to also set such a bar, in order to keep down tech debt, but every known bug should be assessed for severity - which leads to:
> A failing test does not directly benefit the user.
A failing test deemed critical, which prevents a release and keeps the software from destroying hours of a user's work certainly benefits the user, even before any fix is applied!
> End-to-end tests were flaky at times.
Flaky tests are obviously a problem. The article simply asserts that this is a problem for end-to-end tests, and doesn't address why it's not (apparently) a problem for unit tests. I'll try to do so on its behalf, here, as well as indicating where it breaks down: as end-to-end tests involve more complex interactions of more components, we may see a flaky test if any component is variable or if any interaction between components is variable, and in a large enough system we'll see some variability even in the infrastructure it runs on. So I agree that we are more likely to see flaky tests in larger scale tests.
Any test should be designed to be as repeatable as we can make it, and that's usually easier for unit tests. That said, some units may contain race conditions. Where we can prove their absence, we possibly should. But where we can't (and perhaps even where we can) we should still be designing tests to expose these bugs. Those tests will inherently be flaky, so we need some approach to dealing with flaky tests.
It's worth observing that a test that fails one time in a thousand, run ten thousand times, is almost certain (99.995%) to fail at least once. John Hughes talks about doing something like this to successfully test for race conditions. A smart test system could run multifold those tests that are known to be flaky. Further, if we discover a test to be flaky, we could use this same strategy as we walk back the history to discover where the error was introduced.
Flaky tests are indicative of bugs in your tests.
Google is literally advocating treating your bugs as a fact of life, not fixing them, and avoiding harder engineering challenges because they are harder.
This is the worst possible advice.
It can feel a little dangerous, but as long as the integration and e2e tests are kept up and run on build, i think I'm covered. I've definitely been bitten before, but I've been bitten more by over complicated and long running tests.
Here's the thing about "agile" and Unit Tests. The current generation of agile was re-discovered in Smalltalk, after the Smalltalk Refactoring Browser had arrived on the scene. Not only could you refactor at warp speed and high confidence (1) and with unlimited undo, since Smalltalk SUnit test were just ordinary Smalltalk code, Unit tests often got refactored right along with the code.
If you're in an environment where you have to painstakingly refactor, then also separately update your Unit Tests, then it's going to feel like you're doing a whole lot more work -- because you are!
Given that very few environments have the requisite properties, TDD is going to feel like a slog on most stacks, and people will naturally gravitate to something lighter, like integration/acceptance/smoke tests.
(1) - Yes, the prepackaged Refactorings in the original Smalltalk RB were guaranteed to be correct. There were sophisticated runtime checks, and they'd refuse to fire off if correctness wasn't completely certain. You could fool them -- because dynamic language -- but you had to try pretty hard to mess up like that. Then the rest of the shop got really annoyed at you!
I find both Integration and Acceptance Tests could potentially hide many bad internal designs (tight coupling, etc). I also find that not many places can speed up their full integration/acceptance tests suites; suddenly build becomes way too painful, be it local dev build or CI-triggered-through-commit builds.
Note: My experience with super-slow integration tests choking build is with a large JavaScript project with 100 developers (at least).
I implemented integration-tests for a Java-heavy based project before but limited only to test the database component of the project (e.g.: Repository, JPQL, JPA), the ability to test business logics that are tightly coupled with ORM quickly in isolated manner is a blessing.
I guess YMMV.
Integration and functional tests, on the other hand, are long-lived tests intended to ensure that all the pieces stay glued together. They allow the middle parts of code to squiggle quite a bit, as long as they get to the same place at the end of the line. They allow for safe refactoring of modular units whose entire scope lies beneath the boundary of the test. And when the functionality of the application or library changes, they should change as well, but this should be an order or two of magnitude less frequent than the unit test level.
As a rule of thumb, I put integration tests around any piece of functionality that "makes me money"—less literally, around any bit that, if it were broken, would prevent things like user signups, users paying the correct amount, or critical functionality that is indispensible to a user of the software. Likewise, I usually write unit tests when i'm writing new modules so that I don't have to keep reloading a browser or setting up complex data schemes while they're in development. Sometimes I remove those, especially for code that is low-churn (which is a totally under-appreciated metric to look at when you're figuring out where you should focus coverage, IMHO.)
Both types are important necessary. And not all of them are destined to live forever.
testCalculateGrade_scoreOver85_returnsAnA()
Both, unit and integration tests, are imperative. Integration tests point in the general direction of where an issue might be but unit tests are more precise.
I'm not saying that unit tests aren't useful. They definitely help for example if you're using a dynamic programming language or if you're practicing TDD. I would just emphasize a higher priority for integration and acceptance tests.