The Cathedral and the Bazaar, Eric S. Raymond [ebook reader below 3000 txt] 📗

the Linux situation that clearly helps along the Delphi effect is the fact that the contributors for any given project are self-selected. An early respondent pointed out that contributions are received not from a random sample, but from people who are interested enough to use the software, learn about how it works, attempt to find solutions to problems they encounter, and actually produce an apparently reasonable fix. Anyone who passes all these filters is highly likely to have something useful to contribute.

Linus’s Law can be rephrased as “Debugging is parallelizable”. Although debugging requires debuggers to communicate with some coordinating developer, it doesn’t require significant coordination between debuggers. Thus it doesn’t fall prey to the same quadratic complexity and management costs that make adding developers problematic.

In practice, the theoretical loss of efficiency due to duplication of work by debuggers almost never seems to be an issue in the Linux world. One effect of a “release early and often” policy is to minimize such duplication by propagating fed-back fixes quickly [JH].

Brooks (the author of The Mythical Man-Month) even made an off-hand observation related to this: “The total cost of maintaining a widely used program is typically 40 percent or more of the cost of developing it. Surprisingly this cost is strongly affected by the number of users. More users find more bugs.” [emphasis added].

More users find more bugs because adding more users adds more different ways of stressing the program. This effect is amplified when the users are co-developers. Each one approaches the task of bug characterization with a slightly different perceptual set and analytical toolkit, a different angle on the problem. The “Delphi effect” seems to work precisely because of this variation. In the specific context of debugging, the variation also tends to reduce duplication of effort.

So adding more beta-testers may not reduce the complexity of the current “deepest” bug from the developer’s point of view, but it increases the probability that someone’s toolkit will be matched to the problem in such a way that the bug is shallow to that person.

Linus coppers his bets, too. In case there are serious bugs, Linux kernel version are numbered in such a way that potential users can make a choice either to run the last version designated “stable” or to ride the cutting edge and risk bugs in order to get new features. This tactic is not yet systematically imitated by most Linux hackers, but perhaps it should be; the fact that either choice is available makes both more attractive. [HBS]

How Many Eyeballs Tame Complexity

It’s one thing to observe in the large that the bazaar style greatly accelerates debugging and code evolution. It’s another to understand exactly how and why it does so at the micro-level of day-to-day developer and tester behavior. In this section (written three years after the original paper, using insights by developers who read it and re-examined their own behavior) we’ll take a hard look at the actual mechanisms. Nontechnically inclined readers can safely skip to the next section.

One key to understanding is to realize exactly why it is that the kind of bug report non-source-aware users normally turn in tends not to be very useful. Non-source-aware users tend to report only surface symptoms; they take their environment for granted, so they (a) omit critical background data, and (b) seldom include a reliable recipe for reproducing the bug.

The underlying problem here is a mismatch between the tester’s and the developer’s mental models of the program; the tester, on the outside looking in, and the developer on the inside looking out. In closed-source development they’re both stuck in these roles, and tend to talk past each other and find each other deeply frustrating.

Open-source development breaks this bind, making it far easier for tester and developer to develop a shared representation grounded in the actual source code and to communicate effectively about it. Practically, there is a huge difference in leverage for the developer between the kind of bug report that just reports externally-visible symptoms and the kind that hooks directly to the developer’s source-code-based mental representation of the program.

Most bugs, most of the time, are easily nailed given even an incomplete but suggestive characterization of their error conditions at source-code level. When someone among your beta-testers can point out, “there’s a boundary problem in line nnn”, or even just “under conditions X, Y, and Z, this variable rolls over”, a quick look at the offending code often suffices to pin down the exact mode of failure and generate a fix.

Thus, source-code awareness by both parties greatly enhances both good communication and the synergy between what a beta-tester reports and what the core developer(s) know. In turn, this means that the core developers’ time tends to be well conserved, even with many collaborators.

Another characteristic of the open-source method that conserves developer time is the communication structure of typical open-source projects. Above I used the term “core developer”; this reflects a distinction between the project core (typically quite small; a single core developer is common, and one to three is typical) and the project halo of beta-testers and available contributors (which often numbers in the hundreds).

The fundamental problem that traditional software-development organization addresses is Brook’s Law: “Adding more programmers to a late project makes it later.” More generally, Brooks’s Law predicts that the complexity and communication costs of a project rise with the square of the number of developers, while work done only rises linearly.

Brooks’s Law is founded on experience that bugs tend strongly to cluster at the interfaces between code written by different people, and that communications/coordination overhead on a project tends to rise with the number of interfaces between human beings. Thus, problems scale with the number of communications paths between developers, which scales as the square of the humber of developers (more precisely, according to the formula N*(N - 1)/2 where N is the number of developers).

The Brooks’s Law analysis (and the resulting fear of large numbers in development groups) rests on a hidden assummption: that the communications structure of the project is necessarily a complete graph, that everybody talks to everybody else. But on open-source projects, the halo developers work on what are in effect separable parallel subtasks and interact with each other very little; code changes and bug reports stream through the core group, and only within that small core group do we pay the full Brooksian overhead. [SU]

There are are still more reasons that source-code-level bug reporting tends to be very efficient. They center around the fact that a single error can often have multiple possible symptoms, manifesting differently depending on details of the user’s usage pattern and environment. Such errors tend to be exactly the sort of complex and subtle bugs (such as dynamic-memory-management errors or nondeterministic interrupt-window artifacts) that are hardest to reproduce at will or to pin down by static analysis, and which do the most to create long-term problems in software.

A tester who sends in a tentative source-code-level characterization of such a multi-symptom bug (e.g. “It looks to me like there’s a window in the signal handling near line 1250” or “Where are you zeroing that buffer?”) may give a developer, otherwise too close to the code to see it, the critical clue to a half-dozen disparate symptoms. In cases like this, it may be hard or even impossible to know which externally-visible misbehaviour was caused by precisely which bug-but with frequent releases, it’s unnecessary to know. Other collaborators will be likely to find out quickly whether their bug has been fixed or not. In many cases, source-level bug reports will cause misbehaviours to drop out without ever having been attributed to any specific fix.

Complex multi-symptom errors also tend to have multiple trace paths from surface symptoms back to the actual bug. Which of the trace paths a given developer or tester can chase may depend on subtleties of that person’s environment, and may well change in a not obviously deterministic way over time. In effect, each developer and tester samples a semi-random set of the program’s state space when looking for the etiology of a symptom. The more subtle and complex the bug, the less likely that skill will be able to guarantee the relevance of that sample.

For simple and easily reproducible bugs, then, the accent will be on the “semi” rather than the “random”; debugging skill and intimacy with the code and its architecture will matter a lot. But for complex bugs, the accent will be on the “random”. Under these circumstances many people running traces will be much more effective than a few people running traces sequentially-even if the few have a much higher average skill level.

This effect will be greatly amplified if the difficulty of following trace paths from different surface symptoms back to a bug varies significantly in a way that can’t be predicted by looking at the symptoms. A single developer sampling those paths sequentially will be as likely to pick a difficult trace path on the first try as an easy one. On the other hand, suppose many people are trying trace paths in parallel while doing rapid releases. Then it is likely one of them will find the easiest path immediately, and nail the bug in a much shorter time. The project maintainer will see that, ship a new release, and the other people running traces on the same bug will be able to stop before having spent too much time on their more difficult traces [RJ].

When Is a Rose Not a Rose?

Having studied Linus’s behavior and formed a theory about why it was successful, I made a conscious decision to test this theory on my new (admittedly much less complex and ambitious) project.

But the first thing I did was reorganize and simplify popclient a lot. Carl Harris’s implementation was very sound, but exhibited a kind of unnecessary complexity common to many C programmers. He treated the code as central and the data structures as support for the code. As a result, the code was beautiful but the data structure design ad-hoc and rather ugly (at least by the high standards of this veteran LISP hacker).

I had another purpose for rewriting besides improving the code and the data structure design, however. That was to evolve it into something I understood completely. It’s no fun to be responsible for fixing bugs in a program you don’t understand.

For the first month or so, then, I was simply following out the implications of Carl’s basic design. The first serious change I made was to add IMAP support. I did this by reorganizing the protocol machines into a generic driver and three method tables (for POP2, POP3, and IMAP). This and the previous changes illustrate a general principle that’s good for programmers to keep in mind, especially in languages like C that don’t naturally do dynamic typing:

9. Smart data structures and dumb code works a lot better than the other way around.

Brooks, Chapter 9: “Show me your flowchart and conceal your tables, and I shall continue to be mystified. Show me your tables, and I won’t usually need your flowchart; it’ll be obvious.” Allowing for thirty years of terminological/cultural shift, it’s the same point.

At this point (early September 1996, about six weeks from zero) I started thinking that a name change might be in order-after all, it wasn’t just a POP client any more. But I hesitated, because there was as yet nothing genuinely new in the design. My version of popclient had yet to develop an identity of its own.

That changed, radically, when popclient learned how to forward fetched mail to the SMTP port. I’ll get to that in a moment. But first: I said earlier that I’d decided to use this project to test my theory about what Linus Torvalds had done right. How (you may well ask) did I do that? In these ways: