On Sun, Apr 22 2018, Andreas Heiduk wrote: > Am 22.04.2018 um 13:17 schrieb Ævar Arnfjörð Bjarmason: >> >> On Sun, Apr 22 2018, Andreas Heiduk wrote: >> >>> Am 20.04.2018 um 14:14 schrieb Ævar Arnfjörð Bjarmason: >>>> But this is a possible work-around: >>>> >>>> git init /tmp/empty.git >>>> git remote add avar file:///tmp/empty.git >>>> git remote prune avar >>>> git remote remove avar >>> >>> This won't do it also? >>> >>> git remote prune origin >> >> Yes, in this particular case, but that's just emergent behavior in how >> we handle refspec prunign, and the fact that it "works" is arguably a >> bug in "prune". i.e. this: > > Its not emergent because "origin" is the other remote responsible for > that ref and cleaning stuff "belonging" to the remote is the job > description (I'm arguing from a user's perspective, not as a git-developer). Right, I should have said something closer to "inconsistent", at least from the user's perspective. I.e. "remove" doesn't delete your branches because another ref "owns" them, but "prune" will happily clobber another remote's refs without warning. Maybe we're happy to keep that, pruning is a bit of an oddity already, see the "Git has a default disposition of keeping[...]" docs and the rest of the "PRUNING" section I added to git-fetch in 2.17.0, but this is osmething worth keeping in mind. >> >> ( >> rm -rf /tmp/git && >> git clone --bare --mirror git@xxxxxxxxxx:git/git.git /tmp/git && >> cd /tmp/git && >> git remote add avar git@xxxxxxxxxx:avar/git.git && >> git remote add peff git@xxxxxxxxxx:peff/git.git && >> git fetch --all && >> git remote remove avar && >> git remote prune origin >> ) >> >> Will delete all the avar/* and peff/* branches, even though I still have >> a "peff" remote. > > Exactly my point: When you are in a the bad situation of "shared > responsibility", then there is no easy and always correct way out, > because there are uncountable possible situations. > > To give another, slightly modified example expanding the problem space: > > ( > rm -rf /tmp/git && > git clone --bare --mirror https://github.com/git/git.git /tmp/git && > cd /tmp/git && > git remote add github https://github.com/avar/git.git && > git fetch github && > git fetch origin && > # note new fetches for "github/master" using with "(forced update)" > ) > > For ... reasons the first repo publishes some references like > > github/maint > github/master > github/pu > > So when this repo is mirrored AND another, suitably named remote is > added then there will be also namespace conflicts. You can call > > git fetch github > git fetch origin > > in a loop and most likely each fetch will update the same refs, always > toggling between two states. > > So: not only "remote remove" and "remote prune" are at stake but every > command handling remote references. Right, there's certainly a lot of insanity you can create with overlapping refs, but to bring this thread around a bit, the edge cases I'm interested in addressing are those where "git remote <whatever>" silently doesn't do its job, or does it too eagerly, resulting in a hard-to-repair repo state. > How should "git remote remove github" work in both situations? Remove > the refs/remotes/github/master & co? remove them only if the last fetch > was for "github" but not when the last update was for "origin"? Should > "git fetch" use the same logic? Until we move to some other ref store implementation that namespaces things properly, the best we can do is simply to assume that refs/remotes/<name>/ is owned by the <name> remote for the purposes of remove/prune etc. That'll of course leave open this edge case you're pointing out where you're mirroring another remote into refs/* and it creates a remote/<name>/ branch, but I think it's sufficient for the purposes of sane UI to just document that fetching into refs/* creates such caveats. > So it seems better to me to avoid that bad situation altogether. Don't > allow overlapping/conflicting refspecs when adding a new remote. Using > *your* last examples both > >> git remote add avar git@xxxxxxxxxx:avar/git.git && >> git remote add peff git@xxxxxxxxxx:peff/git.git && > > should have failed and hence the "prune" problems won't exist. Same for > my example. I think this is a non-started. There's plenty of legitimate reasons to have overlapping refspecs, e.g. the GitLab case where they're creating a "geo" remote which is a mirror of other refs they push to. Even if that wasn't the case, it would be very fragile to solve these cases by disallowing adding such remotes, since users can edit the config file, we'd need to detect it on the fly. >>> Possible 5): >>> >>> Don't fix "git remote remove" but "git remote add" to complain that its >>> ref-namespace is already occupied by some other remote. Add "--force" >>> for the experts. >> >> Indeed, that's another bug here, i.e. in the above example: >> >> git remote remove peff && # won't delete peff/ branches >> git remote add peff git@xxxxxxxxxx:peff/git.git >> >> Will happily add the "peff" remote again, even though as you point out >> it could be an entirely different remote. > > Ummm. That was not my point. My is: "git clone --mirror" uses a refspec > > fetch = +refs/*:refs/* > > and hence "occupies" the complete "refs/*" namespace. So adding another > remote (for the first time or for the second time is as irrelevant as > the url) will use > > fetch = +refs/heads/*:refs/remotes/peff/* > > and now the "refs/remotes/peff/*" part is in conflict with "refs/*" from > above. The conflict exists not only for "prune" or "remove" but also for > "fetch", "rename" (didn't check) . > > This kind of conflict should not be allowed right from the start - when > the first "git remote add peff..." is executed. Then prune, remove AND > fetch would be OK. As noted above we need to deal with this overlap, and it has to be allowed, but we can still do better than we do now with "remove" / "prune" et al, i.e. simply inform & ask the user what he'd like to do.
