Re: [PATCH v2 02/14] pull: improve default warning
- From
Felipe Contreras <felipe.contreras@gmail.com>
- Date
- Dec 14, 2020, 22:30 UTC
- Message-ID
- <5fd7e77c78e06_d59852083e@natae.notmuch>
- In-Reply-To
- <xmqqh7ooje03.fsf@gitster.c.googlers.com>
Junio C Hamano wrote:
Show 7 quoted lines
> Felipe Contreras <felipe.contreras@gmail.com> writes: > > > This is irrelevant. > > Oh, now you are saying that you do not need a way to squelch these > message to help unconfigured users choose between rebase or merge? > I am confused.
We need a way to squelch the *advise* message, yes.
But this is a secondary priority. The first priority is not to break current behavior users rely on.
Show 9 quoted lines
> > As long as it's an error I don't care if it's short or long. I'm > > against turning on an error from one version to the next. > > Now you are changing your mind? Current version and past ones do > not make it an error to pull non-ff history without choosing rebase > or merge---we go ahead and merge anyway. I thought that in the far > future agreed between two of us, it would be turnend into an error > at some point. There needs one version that turns it into an error > for that to happen. Puzzled.
Yes, in the future, *after* a deprecation period.
Once the users have had a chance to configure git to do what they want, have tried the new mode, haven't had issues with the new mode, or if they disagree with the proposed *future* behavior; complain in the forums at their disposal.
Not before.
Show 32 quoted lines
> >> I too initially thought that pull.mode may be needed, but probably I > >> was wrong. I do think this can be done without pull.mode at all, at > >> least in two ways, without adding different ways to do the same > >> thing. > >> > >> - When pull.rebase is set to 'no' and pull.ff is set to 'only', > >> "git pull" that sees a non-ff history should error out safely. > >> The user is telling that their preference is to merge, but the > >> difference between merge and rebase does not really matter > >> because pull.ff=only would mean we forbid merges of non-ff > >> history anyway. The message you'd get would be "fatal: Not > >> possible to fast-forward, aborting." though. > >> > >> - Or with the advice that hides the latter two points, a user can > >> unset pull.rebase and set the advice.pullNonFF to false to get > >> the same behaviour (i.e. disable the more dangerous half of > >> "pull") with just the "we stopped" error message. > > > > So, after your hypothetical patch, there would be no difference between: > > > > git -c pull.rebase=no -c pull.ff=only pull > > > > and: > > > > git -c advice.pullnonff=false pull > > > > ? > > We do not have to or implement both, but either should give us the > "when pull sees a non-ff history, it should stop without merging or > rebasing, and the user won't be given the advice on how to choose > between merge and rebase" behaviour, I would think.
Right, so both should error out.
And what should these do?
git -c pull.rebase=no -c pull.ff=only pull --merge
git -c advice.pullnonff=false pull --merge
I'm going to answer because I think it's obvious what you would expect: if you pass --merge, both should succeed.
Except they won't, because "git pull --ff-only --merge" fails.
Correct?
Show 8 quoted lines
> >> I think either of these are close enough to what you want, and I > >> think the latter gives us more flexibility in how we tone down the > >> message with advice.pullNonFF. > > > > You are missing at least two things. > > I am guessing that the '?' above I just answered is one you wanted > to ask me, but what's the other one?
Yes. The other is what I explained above; we need a grace deprecation period where we can explain to the user in a simple way what to expect in the future.
And it's much easier to explain to the user that:
git pull --merge -> pull.mode = merge git pull --rebase -> pull.mode = rebase git pull -> pull.mode = fast-forward
Than:
git pull --merge -> pull.rebase = false git pull --rebase -> pull.rebase = true git pull -> pull.rebase = false + pull.ff = only
But those are not the only two. For example there's this additional problem of how to interact with the other values of pull.rebase (other than true and false):
pull.mode = merge pull.rebase = merges
I would expect in this particular configuration that:
git pull -> git pull --merge git pull --rebase -> git pull --rebase=merges
There's no way to represent that with just pull.rebase.
And there's more: some people suggested other modes in 2013, like "pull.mode=none" (essentially "git fetch"), or "pull.mode=merge-reverse-parents".
But the first one should be enough ("git pull --ff-only --merge" doesn't work).
Cheers.
-- Felipe Contreras