Re: [PATCH] config: Introduce GIT_CONFIG_NOGLOBAL

From: "Ævar Arnfjörð Bjarmason" <avarab@gmail.com>
To: Junio C Hamano <gitster@pobox.com>
Cc: Patrick Steinhardt <ps@pks.im>, git@vger.kernel.org
Subject: Re: [PATCH] config: Introduce GIT_CONFIG_NOGLOBAL
Date: Fri, 09 Apr 2021 01:39:54 +0200	[thread overview]
Message-ID: <87wntcgxbp.fsf@evledraar.gmail.com> (raw)
In-Reply-To: <xmqqczv4vgck.fsf@gitster.g>

On Thu, Apr 08 2021, Junio C Hamano wrote:

> Patrick Steinhardt <ps@pks.im> writes:
> [...]
>> Introduce a new GIT_CONFIG_NOGLOBAL envvar, which is the simple
>> equivalent to GIT_CONFIG_NOSYSTEM. If set to true, git will skip reading
>> both `~/.gitconfig` and `$XDG_CONFIG_HOME/git/config`.
>
> I do not think we'd add an unbounded number of new configuration
> sources to the existing three-level hierarchy of system-wide
> (/etc/gitconfig), per-user ($HOME/.gitconfig), and local
> (.git/config), so it is not too bad, from scalability point of view,
> to keep adding new GIT_CONFIG_NO$FROTZ variables.
>
> Let me go in a couple of different tangents a bit, thinking aloud.
>
> Tangent One.  I wonder if the presense of includeIf.<cond>.path
> changes the "we won't have unbounded, so adding another here is OK"
> reasoning.  If somebody wanted to say "Do not use the paths that
> match this and that pattern", it is likely that we'd end up having
> to support a single variable that allows more than one "value".  In
> a straw-man syntax "GIT_CONFIG_EXCLUDE='/home/gitster/*
> /home/peff/*'" might be a way to say "do not use files that are
> under these two directories.
>
> And when that happens, we'd probably notice that it is easier for
> users to configure, if they can treat 'system' and 'global' as just
> another special pattern to place on that list. //system and //global
> might be the syntax we'd choose when time comes, i.e.
>
> 	GIT_CONFIG_EXCLUDE='//system //global'
>
> might become a more scalable replacement for
>
> 	GIT_CONFIG_NOSYSTEM=yes GIT_CONFIG_NOHOME=yes
>
> Tangent Two.  One thing we never managed to properly test in our
> test suite is the unctioning of GIT_CONFIG_NOSYSTEM.  As we do not
> want to get broken by whatever is in /etc/gitconfig, all our tests
> run with the environment variable set.  For the same reason, in
> order to avoid getting influenced by whatever the tester has in
> $HOME/.gitconfig, we export HOME set to the throw-away directory
> created during the test and control what is in the config file in
> that directory.  In hindsight, it might have been a better design to
> use GIT_CONFIG_SYSTEM variable that points at a path to be used as a
> replacement for /etc/gitconfig when it is set; pointing /dev/null
> with the variable would have been the natural way to say "do not use
> anything from system configuration".  And GIT_CONFIG_GLOBAL can
> easily follow the same pattern.
>
> So, from these two perspective, for the longer term end-user
> experience, I am not 100% onboard with GIT_CONFIG_NOGLOBAL.  An
> alternative that allows us to move away from GIT_CONFIG_NOSYSTEM in
> the direction to the tangent #2 would be not to repeat the same
> mistake by doing GIT_CONFIG_NOGLOBAL, and instead adding
> GIT_CONFIG_GLOBAL, which is
>
>  (1) when not set, it does not do anything,
>
>  (2) when set to "/dev/null" (a literal string; you do not have to
>     spell it "NUL" when on Windows), it acts just like the case
>     where your GIT_CONFIG_NOSYSTEM is set,
>
>  (3) when set to any other string, it is taken as a filename that
>      has the global configuration.  Unlike $HOME/.gitconfig or
>      $XDG_HOME/git/config, it is an error if the named file does not
>      exist (this is to catch misconfiguration).
>
> And once this is accepted by users and established as a pattern, we
> could migrate GIT_CONFIG_NOSYSTEM to GIT_CONFIG_SYSTEM=/dev/null
>
>
> Having said all that (meaning: I am not 100% onboard with _NOGLOBAL
> and think _GLOBAL=/dev/null might be a better design), let's give it
> a review under the assumption that _NOGLOBAL is the design we would
> want to choose.

I think doing this via env vars is inherently nasty, but can't think of
a good way to implement it properly without major refactoring.

IMO the "properly" would be that we'd just support this sort of thing as
first-class config syntax, as I've suggested before e.g. in the repo
hooks/config topic.

But to do that we couldn't "stream" the config reading as we do now,
we'd need to read system/global/local, and when we saw some new
meta-syntax apply those ignores to files we already read, and only then
start streaming things to config callbacks.

See "config.ignore.*" in
https://lore.kernel.org/git/87y2ebo16v.fsf@evledraar.gmail.com/