And then you have shocked people when they find out that 100% test coverage doesn't mean that you really have a bug-free codebase.
I understand the point you're making. But, I find this retort to be every bit as frustrating as the rhetoric you're criticizing. It gives us permission to dismiss things we don't fully understand or that require experience and practice to master as snake oil.
Surely most of us would struggle to pick up general relativity or neurosurgery even after weeks or months of training. But we're not--I hope--going to dismiss our neurosurgeon instructor when they say we were cutting something incorrectly even though we REALLY THOUGHT we were doing it right this time.
Maybe TDD really is awesome. And maybe, simultaneously, it's not easy to write a 100% prescriptive guide for how to apply it to every kind of project.
I feel similarly about the rhetoric around OOP. Nine times out of ten, someone complains about OOP and cites some kind of Poodle-Dog-Animal class hierarchy. Then someone comes and says that program object relationships aren't really supposed to be taxonomical. Instead of taking that to heart and wondering if maybe they can try OOP again with a different mindset, the response is defensive. "Surely OOP is still terrible because I was taught the wrong way. And if it's possible to employ a technique ineffectively, then it must be a terrible technique. OOP sucks and they're in a cult so they can't admit it."
Do you have any idea how many wrong ways there are to use the controls in an automobile? Yet we still mostly blame the drivers when they cause a collision because they were using it wrong.
For what it's worth, I think the article is right about the evolution of the term "unit test" and the author mentions the "classicist" approach to unit testing, which is really the one that makes sense with TDD. The "mockist" approach to unit testing, is "simpler" because it just makes an individual class or function the "unit", but that will make tests brittle and make a test-driven approach much more verbose and cumbersome. It also so happens that the mockist approach is the default approach in today's programming languages, testing-frameworks, IDEs, etc.
TDD suffers because it is a tool. As soon as it exists for its own sake, expect that it will be over used. This is more complicated because Tests are themselves a tool. So, having a tool that exists only for the sake of another tool, and it is not too far to see how this one is less clear cut.
I'm not claiming that TDD is good and that it's only the practitioner's fault when things go wrong. Rather, my point is that it's hard to know like the civil engineering example you describe, and that we should be humble enough to acknowledge that "you're doing it wrong" might be a true statement despite how smart we believe ourselves to be.
That is, my point wasn't to say that you can't do it wrong. Rather, doing it right may not further the end goal. Just look up the article of someone trying to TDD a sudoku solver. It is painful, even though there really isn't any one thing that the person did wrong.
I often tell my colleagues that if technologies or methodologies are widely misunderstood, then practitioners aren’t to blame.
TDD might be great, but I have yet to see it widely succeed because its adoption is troublesome. It’s a bit like any of these newest products that promise to address everything consumers demand, only to fail miserably against the same old, leaving the few adopters asking themselves why commoners didn’t get it.
Microservices is another.
TDD gets pushed especially hard (I think) because when it works well it works REALLY well and because it can be quite literally addictive - the red/green being like sounds on a slot machine that generate a hit of dopamine.
Thats a recipe for some passionate promotion.
In fairness to the original author, his heterodox form of TDD does widen its applicability beyond its traditional scope of complex, stateless algorithmic code with simple APIs to the more common integration code that involves databases, etc. that predominates in most commercial code bases.
But humans struggle with rigor, it’s much easier to brand and market something, or to buy in to brands. So “ideas” like microservices become brands. And they’re misunderstood and misapplied because people don’t read the copious literature that discusses the tradeoffs and variations. And they don’t practice it as a discipline with someone that has mastered the technique successfully: they do it blind.
Same goes for TDD or other XP practices that are often deride as cultish. being a discipline is even harder to adopt than a design philosophy like microservices. Disciplines are about consistent behaviour. To an outsider, it’s freaky. But calling it cultish as some do is like saying Karate or another martial art is a cult. From the outside it kind of looks like it, but discipline or kata (practice of form) is known to be a success multiplier for the sustained successful application of practices.
If you don’t have a dojo or a sensei, could you teach yourself such a set of martial arts to mastery? If not, why do we expect everyone to pick up TDD after reading a book?
It makes developers happy because it translates the somewhat arcane nature of the work into something easily digestible by management, and is a fig leaf for shoddy work.
It makes management happy because it goes nicely on a chart that can be shown to the director/customer/shareholders, and looks good at status meetings. It also gives them something to poke at and micromanage.
It makes the customer/shareholders happy because it provides a metric that their money is being spent _doing something_.
TDD may have started as a coding best practice, but it exists and endures - and will continue to exist and endure - because it's performative, and the performance has value to every layer of a business, even though it has nothing to do with actually making the product better at this point.
Get 10 random senior programmers in a room, and see if they agree on matters of design, code structure, testing, frameworks, system architecture. I’ve been part of summits trying to do this several times over the decades, and the only agreement we wound up with was that “Layered abstractions are sometimes a good idea”.
There is a lot of duplication of effort in languages, frameworks, and tools driven by “I can’t be bothered to learn this new idea, I know better”, and a lot of misunderstanding of even first principles.
Something like TDD (and other XP practices like pairing) reminds me of diet/fitness /nutrition regimens that are notoriously difficult to comply with. This isn’t to say they don’t work: athletes and bodybuilders do exist, as do many successful TDD practitioners. It’s just human nature to avoid things that require mental change, or to be prone to self deception in their application.
Its like everything in software engineering: everyone has an opinion, few have actually tested theirs with performant and stable production deployments
Or there's hidden assumptions that differentiate between a Google solution and a coffee shop solution.
Performant... stable... both have ridiculously variable definitions based on who you talk to. That alone is probably enough to get your 10 Sr. developers to provide different solutions.
The counter to TMTOWTDI is a set of principles and discipline, at least within a team or a project. Not to outright reject the alternatives, but to recognize the power of consistent behaviour and prioritization of tradeoffs, to try to counter the brand allegiances.
And yet in the very article we're talking about, there's no explanation of what assumptions the author is operating within.
I'll even go so far to say that most technical articles omit the assumptions - because it's boring content and is less likely to engage the readers.
There are problems where TDD is actually a decent approach. If you have solid requirements beforehand, and your problem does reduce to small bite sized units with trivial state, then it's amazing for producing code that doesn't surprise you when you run it.
There are problems where it really really really isn't. Things where there is ambiguity in the expected outcome. Language processing for example is fairly difficult to tackle with TDD. There is no simple state, there is no unambiguous expected outcome. Whether "jump" is a verb or a noun non-trivially depends on the context, and it may even ambiguous. The correctness of such code is a percentage value, not a boolean.
For context, when I debug the keyword extraction for my search engine crawler, I'm looking at test output that looks like this: <https://memex.marginalia.nu/pics/frog-text.png> (blue are individual keywords, red are potential n-grams).
All that means to me is that when we start a project we think about what it means for an implementation to be correct and write a test. If we agree on the specification, the test, then a passing implementation means that we have implemented it correctly -- for that particular test.
One of the challenges of using unit tests for this kind of testing is, as you say, difficult with modelling continuous values.
It is also difficult because the "proof" is so small and trivial. You need thousands of examples to gain some confidence... and for even moderately complex software this can not be enough.
Unfortunately, the path to writing good tests and verifying software requires a bit of mathematical vocabulary and reasoning which many developers are too shy to learn or outright hostile towards ("I've never needed no stinking math, this is programming!").
My point is: as long as you're driving the development of your software using testing, I think you're practising TDD.
So by "testing their (opinions)", I usually understand people to mean that they've been on a project where they've enjoyed working on something in a particular way (people always underestimate what someone is simply enjoying doing, I have no idea why). I understand that almost nobody is able to compare and contrast two significantly different approaches, let alone establish things like incident rates or cost of improvement and such resulting from application of those approaches (because there's no reference point).
What we do need is for all developers to contribute to setting up a foundation dedicated to establishing software development methodology quality, and then have that set up a dozen teams of 4-8 engineers working on a dozen longish problems (6-36 months) with different methodologies, and then aggregate and analyze results in 12 years. All the engineers would have market-rate salaries, so we'd all have to be chipping in monthly :D
I mean, just imagine where we'd be if we've done that in 2010? (probably at the same place, but just maybe not!)
I see this often when the problem is abstract. When the problem is concrete, I rarely see this come up in practice.
I routinely have 4-5 senior engineers in a room agree on architecture, design and methodology for a single, specific, project. To me, these abstract disagreements are a symptom of a generalization mismatch.
I've found that within the high performing engineers, they tend to agree quite often with the solution once they agree upon the priorities of features and the problems being prevented.
Development practices are not difficult to comply with because you make them work for you, not the other way around.
The issue that I see in software engineering is the XOR approach. Top-down vs bottom up, abstraction vs implementation. This is the wrong thinking. One of the classic books on software architecture said paraphrased "The architecture drives the implementation and the implementation drives the architecture". A classic book on software development said paraphrased "You design for the implementation you need and implement for the design you want.
It's a yin and yang. But many software engineers completely fail at this. They get stuck at implementation or design. You need both at the same time. They're trying to represent a many dimensional problem one or more dimensions lower than required. They're flat-worlders trying to describe a cube or even a tesseract. Of course there's going to be misunderstandings and mistakes. Everyone has a difference view of the cube. They're technically talking about the same thing, yet fail spectacularly when the parts don't quire align when they come together. The problem is very simple in a higher dimension.
But these regimens do work. It's just that they don't work for the reasons most people believe.
Take fasting. Most benefits of fasting in casual practitioners come from the obvious fact that they aren't consuming as much calories as they did before, so they lose weight. That doesn't mean fasting works in a different way from, say, just eating less.
Same goes for TDD. It may be beneficial because it forces a culture of testing and code coverage, but the root of such benefits may not be writing tests before code, yet that's what most people agree TDD is about.
Programming consist of different schools of thought, similar to philosophy or economics, thus conflicting view points are not necessarily wrong per se, but they collide with your school of thought. Therefore there will always be some disagreement even between the most experienced programmers.
Now there are of course still things that a right or wrong on a factual level within programming, but because different schools also labels things as right or wrong it can be hard to distinguish what is grounded philosophical reasons and what is based on factual reasons.
The problem solving aspect of programming is hard for most people. Nearly all of the high performers that I know are on the high functioning autism spectrum. They spearhead the hard technical problems, and set the stage for everyone else to tag along wit the busy work.
It takes all kinds to make a team, but we still need to recognize everyone's strengths and weaknesses.
And not every problem in programming is a technical problem. There are more issues in communication and understanding the problem being solved. Doesn't do much good to have a technically correct product that doesn't quite solve the problem.
What software engineering needs is a methodology in management to identify and properly utilize everyone to their unique abilities.
Take Agile. It clearly outlines that you cannot have managers getting in the way. Yet, I don't know how many companies I've seen trying "Agile" without giving up their managers.
If you're not even willing to move past the first step, indeed it is doomed to failure. But surely you should recognize that when you decide to go down that road?
Removing the managers from the equation requires a team of developers who have a very business-oriented mindset. I expect a lot of developers don't have that, and may not even be capable of it. For any given business, chances are that does not describe your team. I think it is fair to say that Agile is not realistic in the vast majority of cases.
Is the fault Agile for being written for a narrow scope of situations or for practitioners trying to adopt it where it was not designed to fit? I would suggest the latter.
I believe Agile is one of the main offenders here. You suggest that Agile is not "realistic" in most cases, and I agree with that, but Agile consultants and practitioners like to say that Agile is about taking what works for you, and adapting it to your needs.
I think Agile adoption by people who believe in this, is doomed. When Agile is seen as a set of nebulous guidelines from where you can pick and choose, it just doesn't work, and in turn, these experiences feed the idea that Agile doesn't work as a whole.
- We have rearranged your sprints because of super duper important deadline for a very important customer.
- But we already begun this sprint!
- Yeah, but we are Agile.
Compare that to how many of us learned "OOP" or "TDD" or "FP" or "DDD" from a few blog posts.
Is it physicists' fault that I can "learn" that quantum mechanics says CERN opened up a wormhole and alternate timelines from a crystal salesperson on Facebook?
Most of the successful projects, Jay brought in a team of people he had worked with before. So at some point you have to ask if it's the methodology that makes the project successful, or the people who make the methodology work.
If you have ever been a Lead, or even a viable candidate for one, you've had meetings with others to conspire to be successful within the bounds of whatever rules management won't budge on. How to make this number say what we want it to say. How to make that graph go in the right direction. In these cases we are camouflaging our personal methodologies to pass for someone else's. We are succeeding in spite of them, because if we told them what we were doing they might make us stop.
So when they look back at the project, all they see is the things we let them see. The opinion of the person who introduced a change is never the one you can trust. You should ask the people in the trenches what they think.
This applies specially to business people.
100% test coverage just means you tested that s*t with all you expected...
...but your software does not live isolated.
Some external system may have a bug, or may inject some unforeseen values, or a solar flare might flip the value of a bit in your system and crash your software.
fn abs(x: i32) -> u32 {
x as u32
}
fn test_abs() {
assert_eq!(5, abs(5));
}
After discovering the bug, one may even write a branchless implementation of this function for performance without updating the test, and it will still be 100% coverage. But the arithmetic has "logical branches" which do not look like ifs, instead they generate qualitatively different results for different inputs.
In school I wrote parsers. At work I have done it but it's rare.
In fact if you write commercial code you'll often find that the code that you rely on that is like that weirdly gets concentrated in open source libraries which you will be testing only indirectly.
Certainly when writing commercial code I find that the majority of bugs lurk in the interstitial spaces between subsystems I am integrating or in misunderstandings about how the overall system or subparts of it are supposed to behave.
And property tests are not much help there and unit tests are often a hindrance (because theyre as likely to bake in wrong assumptions).
TDD is some help with this but only if A) it's paired with BDD to exorcize specification bugs and B) done with integration tests that exercise all parts of the system together.
College students who start out by writing toy compilers, quicksort implementations and physics engines will often get a distorted view of what professional development is actually like.
If anyone is patronizing here it's you.
I built a property testing library back in the day when I made a library for creating UIs. I was then able to write tests like
* Given an arbitrary UI
* And an arbitrary up/down/left/right list of user direction (this was the only way of navigating)
* If they press a direction and the focus moves, pressing the opposite direction takes them back to where they were before
This uncovered bugs in interfaces like this with 3 items, bottom left is in focus and you press right. Now press left, users probably want to go back to the bottom left rather than top left.
┌─────┐ ┌─────┐ ┌─────┐ ┌──────┐
│ │ │ │ │ │ │ │
│ │ │ │ │ │ │ │
└─────┘ │ │ └─────┘ │ │
│ │ ────► │focus │
┌─────┐ │ │ ┌─────┐ │ │
│focus│ │ │ │ │ │ │
│ │ │ │ │ │ │ │
└─────┘ └─────┘ └─────┘ └──────┘
* Given an arbitrary list of API calls to add/remove/change the UI and user direction presses
* There is always only one item in focus, or no items at all
This actually uncovered a specification bug. We had two requirements
1. Always have an item in focus if there are any that can be in focus
2. If you delete all the items, then add one back in, it's not focussed
Those conflict, but we never noticed, and even had passing unit tests for both cases.
I think property tests can map very nicely to the level of system description that we typically want. I'd love to see larger integration with BDD tools.
wasBusinessDay(date): Date.dayOfWeek(date) not in (Date.Saturday, Date.Sunday)
Both your claim that a big problem is things not working as they are supposed to[1], and the fact that the same old basic security problems get repeated in every new platform that comes along, show that libraries have not, in practice, been any sort of panacea for business software.
[1] I suspect you are referring to application code rather than library code, but either will do for the point I am making here.
You face some hard decisions on that one. None of the error handling paradigms are good at handling the class of numeric errors that people really, really want to pretend don't exist because no matter what kind of error they throw they're a major pain to deal with correctly.
So 100% test coverage you still have tons of possible codepaths that can go sideways, way before including external variability.
And that's why we have code review process, and senior software engineers?
---
EDIT: add an example
When you look at a body of function and think if someone else changes some codes and you will have a hard time figuring out what goes wrong, that, the time you need to add some tests to it. Often it's a function that took numbers of iteration to work as you expect, or a function that is not simple but look so easy you write it right in one go.
And (this gets into personal opinions) sacrificing depth on tests on critical codepaths in order to spend time getting to 100% coverage is also stupid.
Time is finite. Money is finite. 100% coverage provides nothing but false assurances and pretty green metric lights.
It's easy to get "100% code coverage" for the below function:
def nths(number: int, divisor: int) -> float:
return number/divisor
def test_nths():
assert nths(1, 1) == 1
What you need is 100% semantic coverage, i.e. your tests should cover all potential outcomes of the code at appropriate levels.
In many disciplines there are activities that end up feeling like 'scaffolding', learning exercises that you don't necessarily do all the time. They make you a better practitioner the first time you do them, continue to add value for some time, and then are useful from time to time as a refresher.
I think if you've never done TDD, you're missing out. If you had to be dragged kicking and screaming into it, you're missing out (possibly have damaged yourself in the process) You don't need to make a cult out of it (and I suspect some of the resistance to it is from people who feel like they're being asked to join one), anymore than you have to religiously clean your shoes to see value from having a set of cleaning tools.
The same for pair programming. It makes you look at some classes of problem in an new way, and those lessons can stick with you even when you are working alone on a personal project.
Most of the time these days I tend to do a testing sandwich. When I'm still trying to figure out what the APIs will allow I'm free-form writing code trying to stick the concepts together. At some point I get stuck trying to juggle corner cases, and I realize that I'm grinding gears and need to write test code for a while. At this point I still get a lot of the ergonomics benefits of TDD because I'm not yet wedded to my implementation (low sunk cost) - because it doesn't work anyway.
100% code coverage, as other mentioned in this thread, does not guarantee you exercised 100% of all the inputs values possible.
That is why:
- You can pay millions of dollars to a Software Company, and when you install the product the usual license text, says something in capitals along the lines of: NO GUARANTEES WHATSOEVER FOR ANYTHING... plus some will mention you can't use for controlling X-Ray machines, Nuclear Power generators and so.
and also why
- Software where life depends on, normally will use a type of summation or consensus based software like the Shuttle had with 3/4 computers.
If you have a function
F(signed int32 n) -> (signed int32 A, signed int32 B)
A*B = n
It’s next to impossible to have 100% bug free code. Automatic testing helps and TDD helps to write tests, but it’s not the goal. The goal is to create code that is well tested and easy to change.
Not correct. See seL4, CompCert etc.
To use an example like seL4, is formally proven correct against its specification. But how we got the specification itself correct?
"Hidden Fallacies In Formally Verified Systems" https://smartech.gatech.edu/bitstream/handle/1853/62855/BOBE...
How seL4 or CompCert protect me against against hardware issues like ECC errors, or the Pentium FDIV ? [1]
To humbly quote Knuth:
"Beware of bugs in the above code; I have only proved it correct, not tried it."
> How seL4 or CompCert protect me against against hardware issues like ECC errors, or the Pentium FDIV ?
That's a straw man argument. Nobody claims that formal proofs can protect you from hardware failures. However formal proofs can protect you against Pentium FDIV style bugs (see ARM formal specification efforts). Intel started massively investing in formal proof methods because of the not-formally-proven-correct Pentium FDIV fiasco.
In real world use, and other than in academic contexts, is of reduced usefulness that I claim my component used TDD and was formally verified, while throwing over the wall, the responsibility of delivery for a system that can't fail, to somebody else.
So my claim is, we don't know how to do it, we can formally verify some components, and we hope the spec is correct.
Forgetting the spec only has the known known's and unknown known's, but not the unknown unknowns, is a kind of intellectual hubris I am not willing to bet lives on. Certainly not under the disguise of using formal methods.
So I am referring to examples like this:
"A380 Flight Controls overview" [1]
• 3 PRIMary Flight Control and Guidance Computers integration of Auto Flight (ex FGEC) and Flight Control (ex FCPC)
• 3 Auto-Pilots
• 3 SECondary Flight Control Computers dissimilar Software and Hardware, simpler Control Laws
[1] https://www.fzt.haw-hamburg.de/pers/Scholz/dglr/hh/text_2007...
Any technique, in retrospect, could have been applied better, in an improved form.
Traditional TDD proponents would say that what we're doing is definitely not orthodox TDD but IMHO it's the only way that actually works.
I got ya. I have worked on software with a buggered mix of mocked and integration tests that fails about one time in four. But since the bug was in some asynchronous code and would appear in random tests, with random errors (even on thoroughly mocked tests), we can't easily smash it. It's one of those heisenbugs where enabling logging would cause all the tests to pass consistently.
And to put icing this cake, the software worked - this bug wouldn't appear in the production deploy.
So, yeah. The test suite was less than helpful. It certainly couldn't prove any qualities about the software we were writing.