An apology to readers of Test-Driven iOS Development
blog.securemacprogramming.com
blog.securemacprogramming.com
There is a big difference between "unsupported by hard data" and "completely false", especially when the author later admits to not being willing to pay to see the first paper cited in support.
It also seems, (from this post: http://lesswrong.com/lw/9sv/diseased_disciplines_the_strange...) that in fact the first line is supported by data, if only in aggregate. The author of that blog posts makes much of the phrase "this study didn't accurately record" while glossing over the next part of the sentence that states "so we will use average rates summarized from several other studies." It is common practice to aggregate data, and provided it is done properly, the aggregation can reduce error, not increase it.
The claim that "a bug introduced in Requirements, costs more to fix the later it is discovered" is claimed to be supported by data. Well done for demanding to see that data. But epic fail for creating controversy by claiming that the opposite is true.
Redditor 1: "I bet not a single f$%k was given about that table among the readers of the book :)"
Redditor 2: "That table, or something like it, features strongly in every software engineering textbook that I've ever seen. The numbers differ, but the increasing order of magnitude differences are roughly the same. It's the essential justification for nearly everything in software engineering. "A single f$%k" doesn't begin to describe the importance of that table."
And yes, "the essential justification for nearly everything in software engineering" is broadly correct. Anytime you see an argument that "it's important to get the requirements right" for instance, it's based on studies that purport to show these numbers.
The basic falseness of that chart is pretty well known among the hacker set. Or maybe it's a failure of interpretation among the natterers: the "high cost" of mistakes in the past is made mostly of work already spent. The immediate cost of correcting a past mistake is often (of course not always!) vastly lower than people think. And the farther you are from the actual code involved (i.e. if your role is "architect"), the higher your overestimate of that value is going to be.
So this is a feedback loop: the "high level" people at the start of the process have a built-in bias against iterative solutions (which they perceive as "expensive") and thus a built-in predilection to overdesign as they try to get everything right the first time.
You do miss a couple of steps in the feedback loop:
1. "The 'high level' people at the start of the process have a built-in bias against iterative solutions."
2. "A built-in predilection to overdesign as they try to get everything right the first time."
3. The wild-ass guesses that make up the overdesign prove to be completely wrong.
4. The project fails miserably or succeeds, miserably.
5. The "high level" people perceive software as "very expensive", reinforcing their biases.
Well, we've all been there. If you don't get them right, it will really hurt. A lot.
Also, moving things in a wireframe or diagram around is basically free whereas moving things around in a "finished" product can be fairly time consuming.
So, there is at least some truth to it.
Having read this blog post, I now respect him even more. Both for the intellectual honesty and for his efforts to try and reconcile the data in the table, with all its consequences.
I wish more people (including me) were this dedicated and willing to admit mistakes.
Maybe that's what makes him a specialist in software security, of all things.
Surely this makes it more representative, in a literal sense -- if he'd checked his data at the start, he wouldn't have to go back to check it + issue an apology + reprint the books now :P
It's also accepted that this is a very difficult thing to study. There are so many subjective factors involved that's it's really hard to quantify the data or the results. But it seems to me that if Software Engineering as a wants to make progress methodically, it needs to throw out the old unsubstantiated assumptions in order to make room for new conclusions with a solid basis.
I have personally experienced hunting down a bug for six months nearly full time (10 to 12 hours per day) until it was finally found. This was a real-time hardware system and the code was that of an FPGA. The culprit was a coefficient in one of many polyphase finite impulse response filters. The calculations used to generate this coefficient were done in a huge Excel spreadsheet. At one point, in the hundreds of equations in the spreadsheet, the author had used the ROUND() function when the ROUNDUP() function should have been used. This was enough cause a buffer issue in the FIFO feeding the polyphase filter. These are tricky problems to track down when you are literally looking for events that take somewhere in the order of single-digit nanoseconds to occur.
On the other hand, there are those bugs where you know exactly what is going on and where in the code it is happening the instant you see the behavior. We've all experienced that during the course of software development.
Fixing bugs for a dating website has vastly different requirements than, say, fixing bugs in a flight control system.
One argument is for more up-front planning in order to avoid some bugs. At one point this can quickly become counterproductive. Sometimes it's better to just start writing code and fix issues as they come up.
Now, if we are talking about fixing bugs after the fact, that's a different matter. One example here might be if you inherit the code to a complex website or an engine control system and, without any familiarity with the code, are required to fix bugs. This can easily take cubic time as it requires to learn the code-base (and sometimes the subject matter) while also trying to hunt down bugs.
This is why I tend to take such tables or studies with great skepticism. I haven't really paid much attention to these studies, but I remember looking at one or two of them and thinking that they tended to focus on narrow cases such as fixing bugs in-house, with a stable programmer team and great familiarity with the code base.
What is the cost of a bug that is never discovered? Is it negative? What is the cost of fixing a bug that would never have caused a problem (at each stage of development)?
It's way more complex than bugs == bad
Most of these well established claims simply don't have enough empirical data to sustain them. They're all bloggers' hand waving and empty claims.
I am skeptical of this too. It makes more sense to have huge swings in ability, not productivity. Ie a poor programmer won't take 10x as long to code a given feature, he will just hit a ceiling of ability and not be able to do it at all.
10x longer is not a poor programmer, that's an incompetent programmer. If someone needs ten days to code something that can be done in one day something is very seriously wrong.
But given some more nebulous task, where architectural decisions must be made and serious research and testing needs to be done, it's not surprising at all. For example, if you asked someone, "Please write me a library so that I can send and receive XMPP messages," I would expect a large number of otherwise competent programmers to make a significant number of false starts and poor decisions, and generally take much more time than the guy who has experience writing libraries and interpreting text protocols. For example, consider the case of Ron Jeffries and Peter Norvig writing a sudoku solver[1] (this example is a perennial favorite of mine in all sorts of discussions).
And I don't think anything is "very seriously wrong" with this situation. Different skillsets and competence levels result in drastically different results. I think this is true of any profession that is largely about creative problem-solving: some can do it efficiently, some cannot. Programming is just a unique case because there are so many people trying it and not being deterred due to poor performance because there is such a demand.
[1] http://ravimohan.blogspot.com/2007/04/learning-from-sudoku-s...
That's just not true, it's all relative. Linus Torvalds supposedly coded git to the point where it was self-hosted in 1 day. Even a very good programmer could take more than 10 days to do that, an average (but not incompetent) programmer could take months.
http://www.spinics.net/lists/git/msg24132.html
It's still impressive, though.
Edit: Should have read further. It looks like Linus got git self-hosting in two days.
If it encompasses more than just the time of writing code, it becomes even more believable. e.g. A great programmer will take a day to implement the given feature and it will be relatively bug free. An average programmer will take a day to implement the given feature and then it will take another nine to work out the bugs introduced.
Quick, how do you write an SMTP server, and what are the challenges of making it scale?
Most developers won't know how, to start with. That's fine - that's besides the issue. So they need to look it up.
Here starts the performance gap, even if you deal with people with the same lack of knowledge of the relevant RFC's.
In my experience, there's a vast difference in developers ability to read even relatively simple specs and ensure they develop something that follows it. I mentioned SMTP because it genuinely is a simple standard compared to many of the alternatives. But it has enough edge cases that you'll have a big gap out of the gate instantly between the people who have a problem mentally picturing what the spec is describing and those that can easily and systematically map it out.
Secondly in this case you'd start to see experience gaps. Even assuming most people won't have written an SMTP server, you will start seeing a gap between a group of developers that at least have in-depth knowledge of part of the domain or type of service. That will account for a very substantial difference.
In this case, understanding on how to write efficient, scalable network services makes the difference between the guy that will do horribly inefficient stuff like read()'ing a byte at a time to get a line from the client (I mention this because the MySQL C client libraries did that for years instead of the vastly more efficient solution of non-blocking larger reads into a temporary buffer to avoid the context-switches, so it's not like this is something that only rent-a-coder's with no experience will do).
Just the gap between those that understand the tradeoffs of context switches and threads vs. processes vs. multiplexing connections will account for a fairly substantial factor in many types of problems like this.
Then comes the thorny issue of queues. Most otherwise relatively competent developers will struggle to get this right in a way that is not either slow or full of race conditions. Most competent developers never have to deal with really optimizing disk-IO. Witness the wildly different approaches and performance in established mail servers to see that doing queueing well is hard, and those are the good ones.
That does not mean they won't be able to figure out how to do it well enough for typical use cases.
(I used this example, because I've written several SMTP servers, and managed teams working on mail software, so it's an area where I know the tradeoffs and problem points particularly well)
Then again, when writing your typical cookie-cutter web app, the difference probably won't be 10x because so much more of the time will be spent mediating stakeholder requests vs. solving hard problems.
If you compare a developer solving his first task in an unfamiliar language/platform/framework, you will easily see this magnitude of difference compared to a developer with deep experience. But this difference will not stay consistent.
1. Figure out what features should be implemented (ie, will implementing this feature shoot us in the foot later)
2. Figure out the correct implementation
3. Be able to handle future feature requests
Sure, 1 & 2 will vary by skill/experience. However, the skill/experience with which 1 & 2 are handled can severely impact 3, causing it to easily take 10x longer if it can be done at all. As you move onto 4 and down the line, this becomes more and more pronounced.
Things like the Cone, the rising-cost-of-defects or the 10x claim have been kicking around for decades.
The evidence for or against TDD is, admittedly, inconclusive, but it's more recent and of a better academic caliber. There have been a lot of studies. Most of these studies aren't any good - but at least someone is trying.
There's a deeper question, which is "granted that all the empirical evidence we have so far for claims in software engineering isn't all that good, how can we get good empirical evidence?"
I suspect that the answer is going to involve changing the very questions we ask. "Does TDD work" is too fuzzy and ill-defined, and there's no way you can test that in a blinded random experiment. People's biases about TDD (subjects' or experimenters') are going to contaminate the evidence.
Instead, we need to ask questions that aren't susceptible to this kind of bias and contamination. For instance, we might want to unobtrusively study actual programmers working on actual projects, and record what causes them to write defects.
The main problem is that software metrics are imprecise and non-objective. Lines of code, functional points, code coverage, counting code paths, ... we can't have a metric that can be accepted by everyone, all of them have big flaws. And the metrics are the basis for any reliable analysis, if we can't trust them we can't trust anything.
The second main problem is that it is very hard to isolate things under examination. How can we analyse TDD without taking into account the developer's grasp of good design (coupling and cohesion), Dependency Injection and Inversion of Control, refactoring techniques and tools, etc?
Software Engineering is a lot harder because it is much more akin to the fuzzy social studies (e.g.: economics, sociology, management) than to hard sciences (e.g.: computing science).
To address your second point, it's actually not too difficult to isolate factors like TDD. The standard way is to have a control group and a test group. With a large enough of a sample size, you can determine statistical significance with standard tests.
Unfortunately, the test subjects are often university students, who are less experienced than professionals. The fact that data collected on students might not generalize to professionals is a threat to external validity, but should be made explicit in most papers. Most of the time, I think companies aren't very happy about having researchers use their engineers for experiments on the company's dime, but it does happen. So there are some papers out there reporting results with professionals.
Can you point me to some of those studies? Every time I look I only find the same 2 studies everyone quotes from (and aren't very good).
There are 48 papers tagged with "tdd".
Anecdotally, it seems to make sense that it is more expensive to find and fix a defect later in the process. If a dev finds a defect while implementing a feature, only his costs are involved. However, if a defect is found later in the our process, at least 2 qa analysts are involved: one to find the defect and another to confirm it. After that a project manager schedules out time and assigns the defect out to a developer, possibly not the one that introduced the defect gets the assignment to fix it. The developer fixes the defect, a build is made by the build person. The original tester retests the defect and marks the fix as being verified.
That seems complex, but there are possibly even more steps than that. Unit tests may need to be written, customer test cases may need to be updated.
I don't know if the costs end up being exponentially greater, but they would seem to be greater.
At any rate, it would be good to have someone independently validate that data.
I'm not sure what exactly the second QA person adds to the process. As a developer, I need to be able to reproduce a bug myself in order to have any chance of fixing it, so there's your confirmation step right there.
Project manager? No, just have your QA person file the ticket as a bug (with the correct priority) and have your developers pull tickets off your bug tracker themselves. At worst, a slight email/verbal nudge from the usual boss should do.
Build person? No, have a continuous, automated build system. The job of the build person should be to maintain that system, not to run individual builds.
Now you're down to one QA person to file the ticket, one developer to fix the bug, and the same QA person again to close the ticket. Add in a few minutes of another developer's time to code review the first developer (you don't even do that despite having all that process?) and it's still cheaper.
Unit tests? Developers write their own unit tests. Have a failing unit test that reproduces the root cause of the bug before you fix the bug--that's a best practice anyway. For testing above and beyond that, it's not unreasonable to have SDETs maintain that stuff, though you already have the root cause of the bug captured in a unit test so your main concern should be whether any existing tests actually rely upon the broken behavior.
Well - that depends. For plenty of large software engineering projects (think aerospace) it's a given that you'll have that organization many times over.
On the other hand, a web startup may produce a high-quality shipping product with just two or three devops. The number of "lines of code" may even end up being similar between the two deliverables.
The code that gets burned into an ASIC requires a completely different set of development considerations than a non-critical "make deploy_production" web service.
But how do you quantify all these confounding factors?
o The cost to fix the corrupted data
o Opportunity costs of a broken software (it is Black Friday and your site is down)
o Image costs of the ability of your company/product
o The costumer cost (specially if she is also from your company)
I'd always assumed that these numbers referred to an architectural type of bug where, once made, more code is built upon the bug and so a cascading effect occurs. The more code that relies on the bug, the worse it is to fix because you have to fix all of it's dependent code. In some cases you may need to repair data, or entirely refactor sections of an application.
But there's other bugs that are more "typo" level bugs where obviously the time to fix is exactly the same no matter what phase of development.
I had just always assumed those numbers were a worse-case average to encourage developers to better plan their architecture. In part because the context of code complete is more about planning and estimation.
Though there may not be support for the exact figures, I think the point of them is to not let things pile up and to try to put thought into your design especially at the architectural levels.
Making a mistake picking the wrong storage engine can cause major problems if you find it can't scale to your load after 12 months of development. That's a ton of code to change if you're switching from MySQL to Redis (for a crazy example).
Make a mistake in copy or layout, and that's usually much easier to change.
It may be true and noble that since he isn't able to verify the data himself, he shouldn't have used it. But if you can't see the sources, you also can't state that the data is incorrect. One of those books he couldn't find might contain the data that directly corroborates this.
I suspect the cost of fixing defects has more to do with your internal process, and less to do with how much work it takes to find and fix the actual defect. I have a client I work for now for which I need about 10 days lead time to get code from development to production. So it might take 1 hour to fix a bug that is discovered, and 8 hours to do the paperwork and go through the formal process of moving code through testing, staging and finally to production.
The bigger question is whether improvements in processes and tools have obsoleted this data. TDD and automated regression testing would be one place where inter-phase defects could become less costly. On the other hand, projects that have uncorrected architectural errors can be completely functional and yet their maintenance costs never decrease.
If someone has a copy handy, please validate my memory otherwise I'll check when I get home this evening.
Perhaps the closest bit I found was in the chapter "Plan One to Throw Away": "The total cost of maintaining a widely used program is typically 40 percent or more of the cost of developing it" (p 121). Similarly, another section quotes Capers Jones, "Focus on quality, and productivity will follow" (p 217, emphasis original).
That, I think, is about as close as MMM gets to these claims.
The attribution is "Researchers have found" but Boehm and Papaccio are listed in the references.
"The fact that no studies have produced findings that contradict the existence of a teapot orbiting Jupiter..."
On the other hand, what they are trying to study is incredibly complex - not just people, not just software, but people working on software, and even social systems of people working on software. How do you even begin to reliably measure that? Psychology looks like physics by comparison. Yet the resources available to do these studies are a drop in the bucket. This is not something society values – it's not even something our own industry values. So it's unsurprising that what we have so far is at best a smattering. You can argue that these toy studies ought not to be dressed up in formal scientific wear, and I don't disagree – it gives an altogether misleading impression of how solid their conclusions are. But no doubt that's true of a lot of papers.
(Oh and on top of all that, much of the literature is hard to track down. Didn't anyone notice that the OP wasn't able to even find half of the citations he was trying to check? Based on that alone he shouldn't be claiming that "the table is completely false", only that he was unable to rebuild it from data.)
The real question is whether we should throw it all out as junk or try to make something of it. I think there's a good argument for junking. But it's also reasonable to say: hold on, for all their weaknesses these studies are all we have and they're not nothing. At least someone is trying to work with actual data. So let's consider them, but carefully. Like walking on a swamp.
McConnell is one of the few people who have made serious efforts to walk the swamp and filter it somewhat and convey its findings to professional programmers. Of all these people (that I've read), he actually stands out as the most meticulous. To fire cowboy accusations of dishonesty at him personally is likely unfair and misses the point. I do think McConnell overestimates how solid the swamp is, but he's not ignorant about it (read that whole post) and the claim that he outright falsifies citations, or anything like that, demands a high burden of proof. I haven't read Bossavit's critique (should I?) but anyone taking that position had better first make sure he's not standing on a swamp too. That's harder than it sounds.
Author of "Leprechauns" here.
I'm not sure I follow you in the above sentence. What happened time and again in my investigations was that I would find a citation (in McConnell or in Boehm or elsewhere) that was metaphorically accompanied with the statement "this here is a solid piece of land in the swamp".
When I got there, however, what I found was, in fact, just more swamp.
It takes a higher standard of proof to demonstrate that something is solid ground than to demonstrate there's a flaw in it. That may be unfair, but it's why scientists need a lot of training.
To take just one example, Graham mentions the Hughes Aircraft study, cited (by McConnell and others) in support of the usual "exponential" curve for defect cost (as "Willis, Ron R., et al, 1998").
When you actually read that paper - which is both one of the easiest to obtain and one that has the more detailed data about what was studied and how - and look for the raw data, you find numbers that do not obey an exponential law OR a monotonic increase. For one column for instance the numbers in fact vary within a narrow range .36 to 2.00 with two maxima at the "Coding" and "Functional test" phases, dropping off before, between and after. In a different column the costs vary only by a factor of two between Unit Test and System Test, by a factor of less than four between Code and System Test. The exponential rise is generally not true for the pre-1991 period; some post-1991 measurements get closer, but many do not.
It's hard for me to see this evidence as doing anything else but undermining the original claim of the exponential rise as a generally reliable regularity in software development.
Given this, I don't think that citing this paper in support of the claim is appropriate.
I'm saying that if one is going to attack someone personally for dishonesty, as opposed to critiquing the state of the field in general, the bar for that is high and one had better be more than careful with one's own particulars. Reasonable people can interpret this stuff differently. The post linked to upthread [1] didn't strike me as dishonest (though of course it's only one side).
I'm glad you've been looking for solid ground in the swamp and finding it swampy. I think that's valuable. What bothers me are the hints in the OP, comments here, and elsewhere I've seen (including people talking about your work) that this is about one guy making shit up when in reality the problem is endemic to the entire field. Unless he's been egregiously dishonest, which I doubt, it's a distraction.
On another note, you seem like a good person to ask: is there any finding in the software engineering literature that you think holds up? i.e. have you found any solid ground in the swamp? I'm not sure I have (but I haven't looked nearly as hard as you). If there really isn't anything, that alone is kind of shocking.
[1] http://forums.construx.com/blogs/stevemcc/archive/2011/01/09...
Hackers hate ad hominem, and with good reason. I too subscribe to the school of "harsh on the problem, soft on the person". On the other hand, it makes no sense not to call out things that keep us in the swamp, or to tiptoe around important epistemic issues just to spare hurt feelings.
One problem we have is that few people are willing to go to great lengths to check out the available evidence; instead the pattern is to repeat claims (and associated citations) made by people who sound authoritative, accepting them essentially on faith. This has the unfortunate side-effect of magnifying the mistakes of people who have become authorities.
In "Leprechauns" and elsewhere, my focus isn't on what any particular person says, but on specific claims. "The cost of fixing defects rises exponentially as a function of SDLC phase" isn't tied to a particular person - it originated with Boehm but many others have propagated it. My method has been to look up the evidence and to see if it held up. Also to think hard about why studies may have failed to show conclusively what they set out to prove, and how to overcome these challenges.
I'm doing the same kind of thing in my own area, i.e. I'm collecting all available evidence, pro or con, about whether various Agile practices "work".
> is there any finding in the software engineering literature that you think holds up
There are many good ideas and rules of thumb, but when it comes to very general "laws", solidly established - that's harder. I've been asking that question over and over again, hoping to get a convincing answer. Still haven't got one.
I've read a bunch of supposedly solid references, e.g. Pressman, or the "Handbook of Software and Systems Engineering" and have been underwhelmed. (The first "law" proposed in the Handbook: "Requirement deficiences are the prime source of project failures," based on evidence like the Chaos Reports. My rebuttal: http://lesswrong.com/lw/amt/causal_diagrams_and_software_eng... )
There does seem to be a recent wave of software engineering literature, exemplified by http://www.amazon.com/Making-Software-Really-Works-Believe/d... (which I haven't read). Are you familiar with this more recent stuff? Does it represent new research or merely new reporting on old research? If the former, are the standards higher?
You might be able to tell why I lost my enthusiasm for the book.
I think that's the argument for junking the SE literature. If it can't do any better than anecdote, well, to quote Monty Python, we've already got some they're very nice.
What a leap of logic, seriously...
So while the data might still be true, the numbers are not necessarily accurate and thus you shouldn't base conclusions on that table. He doesn't say that the values are wrong - just that they might be incorrect.
On a related note: I'd love to get real numbers for this. Fixing bugs happening in production certainly feels much more expensive and cumbersome, but is there real data aside of this table which now apparently is inaccurate.
"As explained in the book, this table is reproduced from an earlier publication:
Table 1.1, reproduced from Code Complete, 2nd Edition, by Steve McConnell (Microsoft Press, 2004), shows the results of a survey that evaluated the cost of fixing a bug as a function of the time it lay “dormant” in the product. The table shows that fixing bugs at the end of a project is the most expensive way to work, which makes sense…
The data in Table 3-1 shows that, for example, an architecture defect that costs $1000 to fix when the architecture is being created can cost $15,000 to fix during system test.