Slashdot Mirror
Why Is "Design by Contract" Not More Popular?
Coryoth writes "Design by Contract, writing pre- and post-conditions on functions, seemed like straightforward common sense to me. Such conditions, in the form of executable code, not only provide more exacting API documentation, but also provide a test harness. Having easy to write unit tests, that are automatically integrated into the inheritance hierarchy in OO languages, 'just made sense'. However, despite being available (to varying degrees of completeness) for many languages other than Eiffel, including Java, C++, Perl, Python, Ruby, Ada, and even Haskell and Ocaml, the concept has never gained significant traction, particularly in comparison to unit testing frameworks (which DbC complements nicely), and hype like 'Extreme Programming'. So why did Design by Contract fail to take off?"
Comment removed based on user account deletion
Darl put it best: "Contracts are what you use against parties you have relationships with."
I don't have relationships with random other programmers (even if they are female and cute).
GLaDOS for President 2016! "Well here we are again. It's always such a pleasure." -- GLaDOS, 2011
Who cares if it's popular? If it solves your problem, use it.
My other car is first.
Design by contract seems like a lot of extra work and runtime cost for something that might once in a while catch a bug in already-deployed code. Lighter weight methods like static typing catch (certain kinds of) errors before the code is even compiled; unit testing is usually done before code is deployed, and with the express aim of exposing incorrect behavior in corner cases.
Just what niche is design-by-contract supposed to fill? It's heavyweight, costly at runtime, undirected, and likely to catch bugs only after deployment -- too-little-too-late. Maybe it's unpopular because it's a poor tradeoff.
I employ what is known as "Extreme Programming." It mostly involves alot of screaming at the monitor and yelling at my coworkers. I get fired alot but I never have a hard time getting a new job with my Extreme Programming skills that I write all over my resume.
Design by Contract adds more complexity to code, particularly if you're dealing with a language that doesn't natively support it.
Of course, you don't actually need special constructs to check values on input then tossing an exception/returning an error when the data is not in the expected range. In fact, you should go back to remedial programming classes if you're not already doing this.
The last thing you should do is try to figure out what the caller really meant if the value is out of range. Assuming a default works in some cases (which, btw, wouldn't work with Design by Contract as I understand it), but most of the time it's just better to fail and make the programmer fix their mistake.
GLaDOS for President 2016! "Well here we are again. It's always such a pleasure." -- GLaDOS, 2011
If you're serious about design by contract, you need to use a language that supports it. Eiffel does, of course, and so does "Spec#", Microsoft's verifiable variant of C#, but other than that, "support" is a collection of half-baked add-ons that don't provide any strong assurances.
If you're going to take object invariants seriously, you have to take object invariance seriously. Objects can't be allowed to change other than when control is inside them, and when control is inside the object, no public method of the object can be called. This means you have to be able to catch cases where object A calls object B which then calls a public method of A. The invariant of A isn't established at that point, and so, calls into A are illegal. This strict notion of inside/outside is fundamental to class invariants, but many so-called "design by contract" approaches gloss over it. You need a way to explicitly say "control is now leaving this object temporarily" when calling out of an object, and the object's invariant must be true at that exit.
Threading and locking have to be handled in the language. The language needs to know which locks protect what data, or invariants aren't meaningful.
Then there's the problem of how to express an invariant, entry, or exit condition. Are quantifiers provided, or what? How do you talk about concepts like "forward and back pointers of the tree must be consistent"? There's known formalism for that sort of thing, but it's not something you can express cleanly in, say, C or C++.
Without smarts in the compiler, run time checking tends to be too expensive. The compiler needs to know that member function F can't change member variables X and Y, and therefore, invariants concerning X and Y don't have to be rechecked. Without optimizations like that, you end up rechecking everything on every call to every access function.
I'd like to see more design by contract, and I'd like to see it work well enough that when something breaks, you know which side of the interface to blame. I used to do proof of correctness work, and it's quite possible to do this. But you can't do it in C or C++; the languages are too loose. It's been done well for Modula and Java, and a DEC R&D group had a very nice system going just before Compaq canned DEC's Palo Alto research labs. The rise of C killed off verification work; the decline of C may bring it back.
A good business analyst will answer the question "What is it that we (the client) need." They help make sure that the excellent code you write, when it's doing as you were hired to write it to do, solves problems instead of making them.
A good architect will help establish clear separation of authority, giving team members more autonomy to go do what they're good at without having other peoples fingers in their pies or needing to leave their area of scope.
There are a lot of people with pieces of paper from a school that are terrible at these things, and they muddle along leaving wreckage behind them. But that doesn't dismiss the value of having someone competent in those roles when you can find them.
-1 Uncomfortable Truth
You know the aphorism about how any CS problem can be solved by another layer of indirection -- except the problem of too many layers of indirection. That's what design-by-contract is. Instead of having the intrinsic type-safety checks and the social trust that the code author has run unit tests if necessary and makes the code do something reasonable, design-by-contract formalizes all this and makes you specify conditions on the code manually. That's quite a bit of effort for relatively little advantage. The popular design-by-informal-agreement works almost as well and doesn't have the extra, unneeded layer of indirection.
I feel that the reason why design by contract (DBC from now on) isn't popular is because the entire point of the paradigm is that it doubles or triples your code length without adding any actual information; first, you tell the computer what should be true so you can do what you're going to do, then you tell it what to do, then you tell it what you should have done. That's a lot of typing just to make sure the computer fucks up in exactly the way you told it to.
Admittedly, I haven't programmed much in any language that has built in support for DBC, but from exercises in classes (I'm a CS major) I've found that generally it's sort of a waste of time at worst and a duplication of effort at best.
Regardless, the theory remains: if you can write pre- and postconditions for a function, you already know what the function is supposed to be doing so you might as well have spent your time writing the function and doing something else.
For instance, consider some list class's addElement function, with some (sorta) DBC assertions:
(And I apologize for no indenting, but the tabs got stripped out in preview so I'm assuming they're not there when I post)
Of course, this is an overly simple example and I'm probably not even doing it right; however, hopefully it's close enough that you can see what I mean. All of the assertions are semantically redundant; they don't add any meaning to the code. In fact, I don't think it's possible for that to be true in DBC; if an assertion somehow adds information to the code, it's not an assertion any more.
As has been noted, most programmers already do design-by-contract, they just don't call it that. They call it argument checking. The first thing most routines do is validate their arguments, and return an error if any of them are invalid. The last thing done is to check the results and return an error if the results aren't valid. The calling code then checks for error codes or invalid results (eg. a search function returning a null pointer indicating the item wasn't found).
In the real world I often skip this overhead when the conditions are enforced elsewhere. For example, a data structure needed by an internal function may not have to be checked for existence since if it hadn't been created my initialization function would've detected this and signaled an error and the program would've exited. In cases like that, I either omit the check or wrap it in an ifdef so it's only done during development and ignored by the compiler during the release build.
Don't make the mistake of confusing the name of a concept with the concept itself. You'll find quite often that that nifty shiny-new concept someone's presenting as their own has actually been around for 30-40 years and they've just added some chrome, filed off the serial numbers and changed the name to keep you from noticing this.
assert(condition) is your friend. It's not called a contract, it's not design (but a very good practice!) and it does the job well.
I do some teaching on an Open University course here in the UK which uses the concept, and my experience is that many students, including experienced programmers, find it difficult to do. Common errors include:
- confusing the signature of the function (in terms of the types of permitted input) with the pre-condition. It may be true for some implementations of pre-conditions that you need to include information of the form "input is a string" but it isn't for the way we do it in the course.
- ignoring input cases (e.g. giving a post-condition which only makes sense when the input is a non-empty string, but using a "true" pre-condition); students know theoretically that every possible permitted input needs to have an appropriate output in the post-condition but can't put this into practice
- difficulty in creating conditions which are precise; this is both in the early part of the course, which uses English language conditions, and later on when algebraic conditions are introduced
- designing tests which use inputs which fail the pre-condition (which is partly because the testing tool used in the course doesn't check the validity of the pre-condition, so invalid inputs can produce sensible looking outputs)
- confusion between pre- and post- condition rules: they often want to restrict the input by changing the post condition
It seems to be the case that it is at a particular level of abstraction vs practicality that many find difficult to cope with.
You can rent this space for $5 a week.
Finally contracts allow automated testing. That's where you automatically generate data to pass to the code and let the contracts act as a test oracle to catch and locate problems. With something like AutoTest for Eiffel the data generation can be purely random (constrained by preconditions of course), or designed to sample the input according to best coverage via genertic algorithms, etc. The result is that you find corner cases that you might not have anticipated with your unit tests - and you would be surprised how often that happens, AutoTest found a number of subtle bugs in Eiffel's base libraries which had been production code for years. When there's a DbC language or add-on that checks the contracts at compile time, I'll be interested. Then you really need to check out JML and ESC/Java2, and Spec#, because you would be interested.
Craft Beer Programming T-shirts
Apart from "it's too hard", I think Unit Testing has overtaken DbC as an approach.
/ DbcAndTesting.html and particularly the postscript.
- You can write unit tests in any language, with or without a framework. (I saw a "mini-framework" for C that consisted of three macros and a coding convention.)
- In a test, you can specify assertions before and after each method call. It's a little more tedious to represent classic DbC assertions, but the Abstract Test pattern among others allows you to collect common code.
- You can strip out assertions in production code simply by leaving test code out of the product.
- Unit tests also run scenarios automatically, without an extra "test driver".
The one thing Unit Testing *can't* do is check production code as it's running. On the one hand, that's great at catching conditions you never thought of. On the other hand, customers tend to get annoyed if their app shuts down. I'm sure there's some work on partial in the Eiffel world, but so far I haven't seen any.
See also http://onestepback.org/index.cgi/Tech/Programming
They do, yet in my experience, these things are inversely correlated. This applies to all guidance/oversight roles, including business analysts, software architects, consultants, and indeed managers themselves. As I've commented here before, you can always identify a good leader by three characteristics:
The third one is usually the easiest way to identify morons. If you come across a leader in software development who places more value on reports and metrics so they can track things than they do on supporting the developers and test teams working for them, then you know you're dealing with an incompetent.
And yes, that does mean many leaders in software development organisations today are incompetent. That's why the genuinely good people are worth so much.
If you disagree, post your argument. (-1, Overrated) isn't your personal censorship tool for views you don't like.
A lot of functions in real world programs just don't fit this model. Especially in GUIs. Or functions that manipulate internal data structures. Only the most trivial programs (or contrived exercizes of academics) strictly fit the functional, no-side-effects model for all functions. And if you can't apply this method to your entire program, you are going to find a more flexible way to verify behavior.
Those people who think they know everything are a great annoyance to those of us who do. (Isaac Asimov)
I'd agree with much of your post, but I think there's an unwritten assumption about programming style in what you wrote: you seem to be restricting your scope to imperative languages with mutable state (talking about locks and threading, for example).
If you're working in a language that doesn't permit generally mutable state, it's much easier to use concepts of design by contract, essentially because all you have to do is check that when you've finished constructing a new value, it is valid for whatever type it has. Of course, such languages have disadvantages as well.
I suspect that a great deal of work in programming languages over the next few years is going to focus on how to identify and localise side-effects more explicitly. Pure functional languages that don't allow mutable state at all seem to be quite inefficient, and have fundamental problems for high performance applications that have yet to be resolved. Things like the monads widely used in Haskell today provide some powerful features like mutability but built on a much sounder base than many of today's imperative languages, but at the cost of horrendous syntactic overheads, which kinda spoils one of the big advantages of adopting a functional language: conciseness.
However, multi-core and multi-processor machines are fast becoming mainstream, and loose imperative programming languages have failed to provide satisfactory tools to take advantage of these architectures. I expect this to drive a general move towards more declarative rogramming styles in the industry. Meanwhile the academics, who have seen it all before, will be working on more powerful models of scoping and side effects, well beyond the glorified block scope/lambda calculus stuff that most of today's mainstream programming languages are effectively built on. Once we start getting programming languages with more powerful ways to signify when it is acceptable for what sorts of side effects (including changes in state) to occur, we'll have the sort of foundation needed for your ideas about being inside/outside an object, and compilers will have the sort of framework needed to optimise DbC checks so they're only applied when they're really needed and don't carry unfortunate performance penalties.
If you disagree, post your argument. (-1, Overrated) isn't your personal censorship tool for views you don't like.
> assert(condition) is your friend.
And assert(condition && "Explanation of why it's bad and what to do to fix it") is even better. Don't make me read your code and figure out why the hell you put some obscure assert(n != 455) in there.
I especially like contractual gems such as:
char *gets(char *s);
Design by contract, like most formal method approaches, doesn't scale to interesting levels. If you are working on a 200KLOC project on a tight schedule, the last thing you can afford yourself is increasing time spent per line of code by equipping classes, loops and methods with pre & post conditions. And you would need to do this on a substantial scale to make a significant impact on overall quality. I'm sure most projects could boost quality significantly if you double their budgets but then doing so is unacceptable in most real life situations. Good enough involves balancing a lot of factors and quality is just one of them.
It's great if you can specify that a piece of code is a 100% correct implementation of a given specification but in real life the requirements are sketchy at best & keep changing during development. So, you are likely to end up with the wrong system if you don't adjust your interpretation of them to reality during development. Besides, pre and post conditions need maintenance too if you are doing maintenance on your code, so effectively they increase the cost of what is the single most expensive development activity already: maintenance.
Besides there are other, much more useful tools for improving code quality: unit testing, integration testing, static code checkers, compile time type checking, inspections & reviews are all part of the toolkit of an experienced software engineer and largely remove the need for more formal approaches. Additionally clustering and redundant setups are a far cheaper way of guaranteeing uptime than proving the system to be correct. Risk management is better than trying to avoid risk at all cost.
And finally, the value of 100% correctness is overrated. Most commercial software functions acceptably despite the approximately 10 bugs per kloc. In theory disaster could strike any second, in practice it is a rare event that it does and the consequences are quite manageable usually. Of course things do go spectacularly wrong sometimes and usually people then find out a lot was wrong with the overall development process aside from not applying design by contract. So even then, the added value of design by contract is very questionable. You can't compensate for general incompetence with a couple of pre and post conditions.
Jilles
if a test can be expressed as a constraint then it is useful to simply express that as a contract,
Us table-heads who like to shift the processing burden to the database instead of application languages would point out that this resembles database constraints and triggers.
Table-ized A.I.
How interesting for you to put these 2 questions together! Didn't your 2nd question just answered your first question (at least in theory)?
Now in practice, we obvious will be start developing before we got a complete set of requirements, which bring us to...
The big difference between code + unit tests vs contracts (at least from my understanding and experience), is that code and tests are "constructive" in nature, while contracts are mainly "prohibitive" in nature. By "constructive", I mean code and tests tell you something the program will do, vs "prohibitive" which means what the program will not do.
A piece code to parse a string "1234" into an integer 1234 and with test to call it using "1234" and asserting the result == 1234, now that tells you what the code does. The test does not tell you what other things the code might also do, such as it maybe the code can also handle -ve numbers ("-1234"), or formatted numbers ("1,234"), or even decimals "1,234.56" (if the return type is general enough to support it, such as returning type Number in Java).
When you put in assert() into the code, however, it tells what the code does not do. assert(result is integer) tells us the parse does not handle decimals, assert(string contains only numbers) tells us the code does not handles formatted strings, etc.
Comparing the two, code and tests are "constructive" as they give more features to your program, while contracts are "prohibitive" as they restrict what you program can do. More contracts you add, more things you prohibit your code from doing.
When the requirements change ("we have to handle dollar signs in front too, but no more decimals"), the now unused decimal feature and its tests can be ignored, but any contract ("string only contains [0123456789.,-]") that blocks the new requirement has to be removed. Guess which one, tests or contracts, make more problem for the team as the development progress and requirements change?
Oliver.
Unfortunately Mr. Meyer runs up against the halting problem. If the definition of a function, or the contract, says that a parameter must never be the NULL pointer, there are two choices: the code must check whether that pointer is NULL, or it must be proven that that pointer can never possibly be NULL. The second is, with the current state of the art, impossible no matter what language constructs are around to help. That means the check has to be done, the only question is whether it's done by the caller or the called function. And the first rule I learned is to eliminate redundancy, which means that given a choice of doing a check in one place or doing it in a large number of places you do it in the one place.
Design by contract is many things, a large number of them good, but it is not a replacement for error-checking.