As projects get older, they tend to move away from the benevolent dictatorship model and toward more openly democratic systems. This is not necessarily out of dissatisfaction with a particular BD. It's simply that group-based governance is more "evolutionarily stable", to borrow a biological metaphor. Whenever a benevolent dictator steps down, or attempts to spread decision-making responsibility more evenly, it is an opportunity for the group to settle on a new, non-dictatorial system—establish a constitution, as it were. The group may not take this opportunity the first time, or the second, but eventually they will; once they do, the decision is unlikely ever to be reversed. Common sense explains why: if a group of N people were to vest one person with special power, it would mean that N - 1 people were each agreeing to decrease their individual influence. People usually don't want to do that. Even if they did, the resulting dictatorship would still be conditional: the group anointed the BD, clearly the group could depose the BD. Therefore, once a project has moved from leadership by a charismatic individual to a more formal, group-based system, it rarely moves back.
The details of how these systems work vary widely, but there are two common elements: one, the group works by consensus most of the time; two, there is a formal voting mechanism to fall back on when consensus cannot be reached.
Consensus merely means an agreement that everyone is willing to live with. It is not an ambiguous state: a group has reached consensus on a given question when someone proposes that consensus has been reached, and no one contradicts the assertion. The person proposing consensus should, of course, state specifically what the consensus is, and what actions would be taken in consequence of it, if they're not obvious.
Most conversation in a project is on technical topics, such as the right way to fix a certain bug, whether or not to add a feature, how strictly to document interfaces, etc. Consensus-based governance works well because it blends seamlessly with the technical discussion itself. By the end of a discussion, there is often general agreement on what course to take. Someone will usually make a concluding post, which is simultaneously a summary of what has been decided and an implicit proposal of consensus. This provides a last chance for someone else to say "Wait, I didn't agree to that. We need to hash this out some more."
For small, uncontroversial decisions, the proposal of consensus is implicit. For example, when a developer spontaneously commits a bugfix, the commit itself is a proposal of consensus: "I assume we all agree that this bug needs to be fixed, and that this is the way to fix it." Of course, the developer does not actually say that; she just commits the fix, and the others in the project do not bother to state their agreement, because silence is consent. If someone commits a change that turns out not to have consensus, the result is simply for the project to discuss the change as though it had not already been committed. The reason this works is the topic of the next section.
The fact that the project's source code is kept under version control means that most decisions can be easily unmade. The most common way this happens is that someone commits a change mistakenly thinking everyone would be happy with it, only to be met with objections after the fact. It is typical for such objections to start out with an obligatory apology for having missed out on prior discussion, though this may be omitted if the objector finds no record of such a discussion in the mailing list archives. Either way, there is no reason for the tone of the discussion to be different after the change has been committed than before. Any change can be reverted, at least until dependent changes are introduced (i.e., new code that would break if the original change were suddenly removed). The version control system gives the project a way to undo the effects of bad or hasty judgement. This, in turn, frees people to trust their instincts about how much feedback is necessary before doing something.
This also means that the process of establishing consensus need not be very formal. Most projects handle it by feel. Minor changes can go in with no discussion, or with minimal discussion followed by a few nods of agreement. For more significant changes, especially ones with the potential to destabilize a lot of code, people should wait a day or two before assuming there is consensus, the rationale being that no one should be marginalized in an important conversation simply because he didn't check email frequently enough.
Thus, when someone is confident he knows what needs to be done, he should just go ahead and do it. This applies not only to software fixes, but to web site updates, documentation changes, and anything else unlikely to be controversial. Usually there will be only a few instances where an action needs to be undone, and these can be handled on a case-by-case basis. Of course, one shouldn't encourage people to be headstrong. There is still a psychological difference between a decision under discussion and one that has already taken effect, even if it is technically reversible. People always feel that momentum is allied to action, and will be slightly more reluctant to revert a change than to prevent it in the first place. If a developer abuses this fact by committing potentially controversial changes too quickly, however, people can and should complain, and hold that developer to a stricter standard until things improve.
Inevitably, some debates just won't consense. When all other means of breaking a deadlock fail, the solution is to vote. But before a vote can be taken, there must be a clear set of choices on the ballot. Here, again, the normal process of technical discussion blends serendipitously with the project's decision-making procedures. The kinds of questions that come to a vote often involve complex, multifaceted issues. In any such complex discussion, there are usually one or two people playing the role of honest broker: posting periodic summaries of the various arguments and keeping track of where the core points of disagreement (and agreement) lie. These summaries help everyone measure how much progress has been made, and remind everyone of what issues remain to be addressed. Those same summaries can serve as prototypes for a ballot sheet, should a vote become necessary. If the honest brokers have been doing their job well, they will be able to credibly call for a vote when the time comes, and the group will be willing to use a ballot sheet based on their summary of the issues. The brokers themselves may be participants in the debate; it is not necessary for them to remain above the fray, as long as they can understand and fairly represent others' views, and not let their partisan sentiments prevent them from summarizing the state of the debate in a neutral fashion.
The actual content of the ballot is usually not controversial. By the time matters reach a vote, the disagreement has usually boiled down to a few key issues, with recognizable labels and brief descriptions. Occasionally a developer will object to the form of the ballot itself. Sometimes his concern is legitimate, for example, that an important choice was left off or not described accurately. But other times a developer may be merely trying to stave off the inevitable, perhaps knowing that the vote probably won't go his way. See “Difficult People” in Kapitel 6, Communications for how to deal with this sort of obstructionism.
Remember to specify the voting system, as there are many different kinds, and people might make wrong assumptions about which procedure is being used. A good choice in most cases is approval voting, whereby each voter can vote for as many of the choices on the ballot as he likes. Approval voting is simple to explain and to count, and unlike some other methods, it only involves one round of voting. See http://en.wikipedia.org/wiki/Voting_system#List_of_systems for more details about approval voting and other voting systems, but try to avoid getting into a long debate about which voting system to use (because, of course, you will then find yourself in a debate about which voting system to use to decide the voting system!). One reason approval voting is a good choice is that it's very hard for anyone to object to—it's about as fair as a voting system can be.
Finally, conduct votes in public. There is no need for secrecy or anonymity in a vote on matters that have been debated publicly anyway. Have each participant post her votes to the project mailing list, so that any observer can tally and check the results for herself, and so that everything is recorded in the archives.
The hardest thing about voting is determining when to do it. In general, taking a vote should be very rare—a last resort for when all other options have failed. Don't think of voting as a great way to resolve debates. It isn't. It ends discussion, and thereby ends creative thinking about the problem. As long as discussion continues, there is the possibility that someone will come up with a new solution everyone likes. This happens surprisingly often: a lively debate can produce a new way of thinking about the problem, and lead to a proposal that eventually satisfies everyone. Even when no new proposal arises, it's still usually better to broker a compromise than to hold a vote. After a compromise, everyone is a little bit unhappy, whereas after a vote, some people are unhappy while others are happy. From a political standpoint, the former situation is preferable: at least each person can feel he extracted a price for his unhappiness. He may be dissatisfied, but so is everyone else.
Voting's main advantage is that it finally settles a question so everyone can move on. But it settles it by a head count, instead of by rational dialogue leading everyone to the same conclusion. The more experienced people are with open source projects, the less eager I find them to be to settle questions by vote. Instead they will try to explore previously unconsidered solutions, or compromise more severely than they'd originally planned. Various techniques are available to prevent a premature vote. The most obvious is simply to say "I don't think we're ready for a vote yet," and explain why not. Another is to ask for an informal (non-binding) show of hands. If the response clearly tends toward one side or another, this will make some people suddenly more willing to compromise, obviating the need for a formal vote. But the most effective way is simply to offer a new solution, or a new viewpoint on an old suggestion, so that people re-engage with the issues instead of merely repeating the same arguments.
In certain rare cases, everyone may agree that all the compromise solutions are worse than any of the non-compromise ones. When that happens, voting is less objectionable, both because it is more likely to lead to a superior solution and because people will not be overly unhappy no matter how it turns out. Even then, the vote should not be rushed. The discussion leading up to a vote is what educates the electorate, so stopping that discussion early can lower the quality of the result.
(Note that this advice to be reluctant to call votes does not apply to the change-inclusion voting described in “Stabilizing a Release” in Kapitel 7, Packaging, Releasing, and Daily Development. There, voting is more of a communications mechanism, a means of registering one's involvement in the change review process so that everyone can tell how much review a given change has received.)
Having a voting system raises the question of electorate: who gets to vote? This has the potential to be a sensitive issue, because it forces the project to officially recognize some people as being more involved, or as having better judgement, than others.
The best solution is to simply take an existing distinction, commit access, and attach voting privileges to it. In projects that offer both full and partial commit access, the question of whether partial committers can vote largely depends on the process by which partial commit access is granted. If the project hands it out liberally, for example as a way of maintaining many third-party contributed tools in the repository, then it should be made clear that partial commit access is really just about committing, not voting. The reverse implication naturally holds as well: since full committers will have voting privileges, they must be chosen not only as programmers, but as members of the electorate. If someone shows disruptive or obstructionist tendencies on the mailing list, the group should be very cautious about making him a committer, even if the person is technically skilled.
The voting system itself should be used to choose new committers, both full and partial. But here is one of the rare instances where secrecy is appropriate. You can't have votes about potential committers posted to a public mailing list, because the candidate's feelings (and reputation) could be hurt. Instead, the usual way is that an existing committer posts to a private mailing list consisting only of the other committers, proposing that someone be granted commit access. The other committers speak their minds freely, knowing the discussion is private. Often there will be no disagreement, and therefore no vote necessary. After waiting a few days to make sure every committer has had a chance to respond, the proposer mails the candidate and offers him commit access. If there is disagreement, discussion ensues as for any other question, possibly resulting in a vote. For this process to be open and frank, the mere fact that the discussion is taking place at all should be secret. If the person under consideration knew it was going on, and then were never offered commit access, he could conclude that he had lost the vote, and would likely feel hurt. Of course, if someone explicitly asks for commit access, then there is no choice but to consider the proposal and explicitly accept or reject him. If the latter, then it should be done as politely as possible, with a clear explanation: "We liked your patches, but haven't seen enough of them yet," or "We appreciate all your patches, but they required considerable adjustments before they could be applied, so we don't feel comfortable giving you commit access yet. We hope that this will change over time, though." Remember, what you're saying could come as a blow, depending on the person's level of confidence. Try to see it from their point of view as you write the mail.
Because adding a new committer is more consequential than most other one-time decisions, some projects have special requirements for the vote. For example, they may require that the proposal receive at least n positive votes and no negative votes, or that a supermajority vote in favor. The exact parameters are not important; the main idea is to get the group to be careful about adding new committers. Similar, or even stricter, special requirements can apply to votes to remove a committer, though hopefully that will never be necessary. See “Committers” in Kapitel 8, Managing Volunteers for more on the non-voting aspects of adding and removing committers.
For certain kinds of votes, it may be useful to expand the electorate. For example, if the developers simply can't figure out whether a given interface choice matches the way people actually use the software, one solution is to ask to all the subscribers of the project's mailing lists to vote. These are really polls rather than votes, but the developers may choose to treat the result as binding. As with any poll, be sure to make it clear to the participants that there's a write-in option: if someone thinks of a better option not offered in the poll questions, her response may turn out to be the most important result of the poll.
Some projects allow a special kind of vote known as a veto. A veto is a way for a developer to put a halt to a hasty or ill-considered change, at least long enough for everyone to discuss it more. Think of a veto as somewhere between a very strong objection and a filibuster. Its exact meaning varies from one project to another. Some projects make it very difficult to override a veto; others allow them to be overridden by regular majority vote, perhaps after an enforced delay for more discussion. Any veto should be accompanied by a thorough explanation; a veto without such an explanation should be considered invalid on arrival.
With vetoes comes the problem of veto abuse. Sometimes developers are too eager to raise the stakes by casting a veto, when really all that was called for was more discussion. You can prevent veto abuse by being very reluctant to use vetoes yourself, and by gently calling it out when someone else uses her veto too often. If necessary, you can also remind the group that vetoes are binding for only as long as the group agrees they are—after all, if a clear majority of developers wants X, then X is going to happen one way or another. Either the vetoing developer will back down, or the group will decide to weaken the meaning of a veto.
You may see people write "-1" to express a veto. This usage comes from the Apache Software Foundation, which has a highly structured voting and veto process, described at http://www.apache.org/foundation/voting.html. The Apache standards have spread to other projects, and you will see their conventions used to varying degrees in a lot of places in the open source world. Technically, "-1" does not always indicate a formal veto even according to the Apache standards, but informally it is usually taken to mean a veto, or at least a very strong objection.
Like votes, vetoes can apply retroactively. It's not okay to object to a veto on the grounds that the change in question has already been committed, or the action taken (unless it's something irrevocable, like putting out a press release). On the other hand, a veto that arrives weeks or months late isn't likely to be taken very seriously, nor should it be.