Re: [PATCH] push: allow atomic flag via configuration

From: Jonathan Nieder <jrnieder@gmail.com>
Date: Mar 28, 2017, 00:51 UTC
Message-ID: <20170328005126.GP31294@aiede.mtv.corp.google.com>
In-Reply-To: <1490603272.9788.7.camel@locke.gandi.net>

Hi,

Romuald Brunet wrote:

> On ven., 2017-03-24 at 12:29 -0700, Junio C Hamano wrote:
>> Jeff King <peff@peff.net> writes:

Show 24 quoted lines

>>> My one question would be whether people would want this to actually be
>>> specific to a particular remote, and not just on for a given repository
>>> (your "site-specific" in the description made me think of that). In that
>>> case it would be better as part of the remote.* config.
>>
>> Yeah, I had the same reaction.
>>
>> Conceptually, this sits next to remote.*.push that defines which set
>> of refs are sent by default, and remote.<name>.pushAtomic does make
>> sense.  If (and only if) it turns out to be cumbersome for somebody
>> to set the configuration for each and every remote, it is OK to also
>> add push.atomic to serve as a fallback for remote.*.pushAtomic, I
>> would think, but adding only push.atomic feels somewhat backwards.
>
> Thanks for your feedback
>
> I'm mostly using single remotes that's why I didn't even think of making
> it configurable per remote. But you're right that makes more sense.
>
> I'll try to make that modification to the patch.
>
> As for my use case: I'd like to use default atomic pushes when pushing a
> new tag among our stable branch, but inevitably forgetting to rebase
> beforehand. Therefore pushing a "dangling" commit/tag

Making it per-remote would also address my concern about scripts. Existing scripts may be using

	git push $url $refs

and relying on the push to partially succeed even when some refs cannot be fast-forwarded. (For example, I had such a script that did

	git push my-mirror refs/remotes/origin/*:refs/heads/*
	git push my-mirror +origin/pu:pu

because the first command would fail to update pu, and then the second command would clean up after it.) A configuration that globally changes the effect of "git push" to mean "git push --atomic" would break such scripts. A per-remote configuration is tightly enough scoped to not be likely to cause unrelated breakage for users that run scripts written by others.

Thanks and hope that helps, Jonathan