On Mon, Feb 6, 2012 at 18:14, Jakub Narebski <jnareb@xxxxxxxxx> wrote: > On Mon, 6 Feb 2012, Johan Herland wrote: >> On Mon, Feb 6, 2012 at 15:44, Jakub Narebski <jnareb@xxxxxxxxx> wrote: >> > Relying on (default) hooks to implement this feature has the disadvantage >> > that it wouldn't be turned on by default... while this feature would be >> > most helpful for users new to git (scared by refuse to push). >> >> True. I too believe that this will be most helpful if it is enabled by >> default. That said, the easiest way to get there might be through >> first demonstrating that it works in practice when implemented as >> hooks. > > Yes, starting with prototype implementation using existing infrastructure > (hooks) would be a very good idea. (That's how first versions of what > became submodules were implemented.) > > OTOH we should be aware of limitations of said prototype due to the fact > that it is a prototype... Agreed, but AFAICS (and modulo the addition of pre-rewrite and pre/post-push hooks mentioned earlier) all of the things discussed so far in this thread can be implemented as hooks. >> > I am not sure either if everything (wrt. safety net) can be implemented >> > via hooks. One thing that I forgot about is preventing rewinding of >> > branch past the published commit using e.g. "git reset --hard <commit>". >> > Unless `pre-rewrite` hook could be used for that safety too... >> >> Hmm. I don't think we'll be able to "plug" all the holes that might >> leave the user in a rewritten state (e.g. what if the user (possibly >> with the help of some tool) does an "echo $SHA1 > >> .git/refs/head/master"?). > > First, I was thinking about having safety net against rewriting published > commits being present only in porcelain. Plumbing would be not affected > (perhaps there would be need to extend or add new plumbing to query "phase" > state, though). I also think this is very much a porcelain-only feature. I'd be more worried if plumbing changes were needed. >> And trying to plug too many holes might end >> up annoying more experienced users who "know what they're doing". > > Second, forcing via command line parameter should always be an option. Obviously, but if a large portion of the Git community felt the need to always disable this feature, I'd say that we'd failed. The best features are those that Just Work(tm). >> Instead we might want to add a client-side check at push time. I >> realize that this check is already done by the remote end, but the >> client-side might be able to give a more helpful response along the >> lines of: > > [...] > > Explanation is good, but the whole idea of rewriting safety is that you > are informed (warned or denied) _before_ attempting rewrite and doing much > work. True, but there may be cases where the rewrite is not apparent until after it has happened. E.g. a novice user may use 'git reset --hard' in order to get to an earlier state for testing purposes, and then - after completing the test - 'git reset --hard' back to the starting point. I know this is not best practice, but is it bad enough that we want to refuse it? >> First, we don't need to annotate _all_ commits. For the 'public' >> state, we only annotate the last/tip commit that was pushed/published. >> From there, we can defer that all ancestor commits are also 'public'. > > Right. > >> For the 'secret' state, we do indeed annotate _all_ secret commits, >> but I believe this will be a somewhat limited number of commits. If >> your workflow forces you to annotate millions of commits as 'secret', >> I claim there is something wrong with your workflow. > > Well, for the 'secret' we can rely on the fact that child of 'secret' > commit must also be 'secret' (non-publishable) if secret is to stay > secret. Still marking all 'secret' commits might be better idea from > UI and from performance point of view. I don't think we should automatically assume that all children of a 'secret' commit are also 'secret'. First of all, the git DAG was not made for iterating forwards in history, so given a 'secret' commit, it is computationally expensive to enumerate all its implied-'secret' descendants. More importantly though, I don't agree with the premise. I would typically use the 'secret' state as follows: While debugging a piece of code, I might commit a few debug print statements, and I would typically mark this debug commit as 'secret', in order to prevent myself from accidentally pushing this. Although it probably doesn't matter in practice, I think it is wrong for the commits made (temporarily) on top of this debug commit to be also considered 'secret'. They are only unpublishable as a consequence of being based on the debug commit, and only until I get around to rebasing away the 'secret' debug commit. >> > Second, I have doubts if "phase" is really state of an individual commit, >> > and not the feature of revision walking. > > It matters to presentation: can commit be simultaneously 'public' because > of one branch, and 'draft' because of other. > >> I believe the 'public' state is a "feature of revision walking" (i.e. >> one annotated 'public' commit implies that all its ancestors are also >> 'public'). However, the 'secret' state should be bound to the >> individual commit, IMHO. > > Good call, otherwise 'secret' commit could have been "side-leaked" > by other refs being pushed. > > This means though that 'public' / 'draft' while looking similar to 'secret' > are in fact a bit different things. In other words 'immutable' and > 'impushable' traits are quite a bit different in behavior... > > Especially that one acts at pre-rewrite time, and second pre-push time. Exactly. I find Mercurial 'phase' language confusing, precisely for the reason that 'public' and 'secret' are DIFFERENT concepts. One hinders rewrite and naturally applies to a commit AND its ancestors, while the other hinders push and only applies to the commit itself. The fact that they could be implemented by the same mechanisms (hooks and notes) does not make them the same thing. >> > Take for example the situation where given commit is reference by >> > remote-tracking branch 'public/foo', and also by two local branches: >> > 'foo' with upstream 'public/foo', and local branch 'bar' with no upstream. >> > >> > Now it is quite obvious that this feature should prevent rewriting 'foo' >> > branch, for which commits are published upstream. But what about branch >> > 'bar'? Should we prevent rewriting (e.g. rebase) here too? What about >> > rewinding 'bar' to point somewhere else. What if 'bar' is really detached >> > HEAD? >> > >> > These questions need to be answered... >> >> Good point. There are two questions we may need to answer: "Has commit >> X ever been published?", and "Has commit X ever been published in the >> context of branch Y?". In the latter case, we do indeed need to take >> the upstream branch into account. > > I think the second one is more interested for rewrite safeties. > >> Basically, there are three different "levels" for this rewrite/publish >> protection to run at: >> >> 1. Do not meddle at all. This is the current behavior, and assumes >> that if the user rewrites and pushes something, the user knows what >> he/she is doing, and Git should not meddle (obviously unless the >> server refuses the push). > > I think that there should be some easy way to force such behavior, > i.e. to discard rewrite safeties. Indeed. We should probably have a simple config flag to enable the rewrite protection. In fact I would argue that the flag should default to false (disable protection) when unset, and then we should let init/clone set the config flag to true (enable protection) in newly created repos (unless explicitly disabled in the user/system configs). This way, behavior does not change for existing repos, but new repos are protected by default (with only a single command needed to disable the protection). >> 2. Warn/refuse rewriting commits in your upstream. This would only >> check branch X against its registered upstream. Only if there is a >> registered upstream, and you're about to rewrite commits that are >> reachable from the upstream remote-tracking branch, should Git >> intervene and warn/refuse the rewrite. This level would IMHO provide >> most of the benefit, and little or no trouble (i.e. false positives). > > Right. I wonder if we can get usage statistics from Mercurial users > about usage of their "phases" feature... though mapping terminology > for example 'upstream' from Git to Mercurial and vice versa can be > a pain, I guess. I'm unsure how useful it would be. IMHO Git and Mercurial are different enough (and promote sightly different workflows) that I don't trust the the average Mercurial user's preference for Mercurial's 'phase' behavior to be transferable to the average Git user's preference for a similar behavior in Git. >> 3. Warn/refuse rewriting _any_ 'public' commit. Refuse to rewrite any >> commit that is reachable from any remote-tracking branch. Some would >> say that this is a Good Thing(tm), since it prevents a commit from >> being _copied_ (i.e. rebased or cherry-picked) between branches (you'd >> be in this camp if you run a tightly-controlled workflow, where you >> e.g. mandate upmerging patches from the oldest applicable branch >> instead of cherry-picking patches from a newer branch). However, other >> people would say that this is too limiting, and imposes unnecessary >> rules on the workflow of the project (where e.g. copying (by way of >> git-rebase) a topic branch from one place to another would cause an >> annoying false positive). > > Well, we could always 'deny' on 2nd, and just 'warn' on 3rd... Good idea. Have fun! :) ...Johan -- Johan Herland, <johan@xxxxxxxxxxx> www.herland.net -- To unsubscribe from this list: send the line "unsubscribe git" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html