Ideally, it's an emotionless affair. The community just repels those who do not work well inside of it. The idea is just to get them to exercise their RightToLeave, no more and no less. Later on, if the individual in question decides to return to try again, the community is willing to accept that, a la ForgiveAndForget, because it doesn't really matter. The community is secure enough in itself that it does not have to perceive any one individual as a threat.
There are various ways to persuade people to leave: DissuadeInteraction, DissuadeReputation, PunishReputation, CommunityWarning, PeerPressure, etc. Common solutions are to block the IP or login of the attacker (hard), or to delete their content (soft), or AutomaticSandbox (a combination of the two). These methods can be sufficient. On the other hand, they can also create deep animosity resulting in real attacks on a system. This was the case with the spam attack on KuroShin in Summer 2000, where PeerPressure created an attacker rather than an exile. Tough problem. Maybe the real solution is to try RadicalInclusiveness instead?
In theory, exiling is HardSecurity: it creates an inflexible dichotomy between those who are part of the community, and those who are not: SoftSecurity prefers fuzzier boundaries. In practice, though, exiling is, at least in part. SoftSecurity. People are typically exiled slowly, over the course of days, weeks, or even months. An exile may contact members of the community through email or by some other method. She may continue to read what is written. She may continue to have an effect through things she wrote prior to her exile, or things she writes on other sites. She may quietly circumvent the ban, as discussed below. Realistically, she can continue to hover round the edges of the community long after the bulk of its members have forgotten she ever existed, which is pretty scary.
On most online communities, "bans" or "suspensions" can be circumvented. Technological measures are typically quite weak: users can obtain new IP addresses, new logins, and so forth, and thus "reincarnate" no matter how many times they are slain. People who persistently refuse to be exiled come to be viewed as phoenixes, forever rising from their ashes, or hydras, growing two new SockPuppet identities every time one is cut off.
Such reincarnations can either be stealthy, or blatant: that is, an exiled user may return in a way that clearly signals that they were once banned, or they may try to blend back in without being noticed. Provided that there's a smidgeon of FairProcess in the decision to exile someone, blatant reincarnation is not a serious threat to the community, as it's a fairly obvious mark of trolldom (see WhatIsaTroll). This is the basis of the TrialByExile pattern.
In many ways, stealthy reincarnation is productive, as it acts as a form of DissuadeReputation - in order to avoid being exposed as a reincarnation, the exile must act in a different way, talk to different people, and so forth. Thus, a subverted exile can still be an instance of transformative justice. However, it raises some issues:
Very short "suspensions" are used to implement TrialByExile.
Always assign a time limit, say a month or a year, after which the exilee is free to return. This mimics the criminal correctional systems we have in wider society. LifetimeBans are bad.
A stronger form of CommunityExile is an EnforcedExile?, where someone has been prevented from having any effect on a community. Examples of an enforced exile would include deportation from a country or being banned from a club. In some ways, you might even consider imprisonment a form of exile, although that isn't really true as the society still takes some responsibility for the person. Here, the basic idea is to physically place the person outside of the geographic or legal boundaries of the society.
This is essentially only possible where you can create high barriers to entry, such as a GatedCommunity. You need some sort of authentication mapping to the RealWorld to do this for real, such as a TrustMetric or an GovernmentBackedAuthentication. This HardSecurity form of CommunityExile cannot be used in most online communities, which deliberately have low barriers to entry (LoginsAreEvil).
Unlike a ban in software which is encoded in an unforgetting database, when a person is exiled psychologically (i.e. through social means), this ban has to be propogated through the memories and intentions of the community. Since the community changes, and newcomers have NoRespectForHistory of fights, they will tend to ForgiveAndForget the old situation simply because they never cared in the first place. Think of this as a DynamicValue in wetware. Old timers may have a hard time forgiving and forgetting, but they should try to ForgiveAndForgetInWetware.
If the old timers are trying to ForgiveAndForget the argument, they may never reach catharsis because they cannot talk about the situation freely. Even if the exilee continues to do bad things, the community is expected to keep quiet, even if this will further ennervate old wounds. This may push the conflict further down the conflict spiral rather than alleviating it as was the desired effect.
Failing to ForgiveAndForget may make the old timers seem irrational, necessarily psychotic (qua psychological ban), to newcomers who don't understand the old conflict. This is ironic since if the ban was in software when the community did agree, the newcomers would likely accept the ban as legitimate. The constant political arguing however makes the ban enlivened and real to the newcomers, rather than something they could ignore, and it will likely encourage the old timers to harden their positions rather than get around to the difficult task of forgiving and forgetting.
Therefore, record a short case history of the psychological ban and then assign a time limit, say a month or a year, after which the exilee is free to return. This mimics the criminal correctional systems we have in wider society. Don't forget to delete the short case history, or sufficiently anonymize it. You may actually want to pseudonymize the case history initially, placing it in a seperate ExileLog, as long as the community remembers which case history belongs to whom; this will prevent leaving a monument to the exilee's poor reputation (cf. PunishReputation).
If your community uses technology exiles, then you might think this section is irrelevant to you. However, as noted above, HardSecurity exiles are practically impossible - you can always get round a block if you're motivated enough. EnglishWikipedia is no less psychotic than MeatballWiki over this, so CryptoNauts needn't gloat.