Re: [Outreachy PATCH v2] environment: move "core.attributesFile" into repo-setting

[Phillip Wood <phillip.wood123@gmail.com>]

On 05/01/2026 22:24, Junio C Hamano wrote:
> Phillip Wood <phillip.wood123@gmail.com> writes:
> 
>> If I run 'git -c core.attributesFile=~does-not-exist rebase -i' with git
>> built from master it fails immediately with "fatal: failed to expand
>> user dir in: '~does-not-exist'".
> 
> Hmph, if you call any behaviour change a "regression", this may
> certainly count as one, but I do not necessarily think the above is
> a good behaviour.

One can argue that it depends on the command, but a as far as "git 
rebase" is concerned this change in behavior is a regression. It is not 
just rebase that is affected, for example "git merge" in a partial clone 
now downloads all the blobs it wants before erroring out which isn't 
terrible but it is hard to argue that's an improvement. I suspect "git 
cherry-pick" and "git revert" are also negatively impacted by this change.

> Think about a use case where attributes are not used at all, e.g.,
> "git -c core.attributesFile=~does-not-matter cat-file -t HEAD";
> would it make sense to barf when your configuration file has an
> invalid definition for what you are *not* using?  So if the change
> makes it stop barfing, it can even be argued that this is an
> improvement.

For some commands, but the fact that other commands now die when they're 
not expecting to and so end up in a strange state is a regression. Given 
how widespread the use of attributes is it would be hard to audit all 
the affected code paths and adjust them to the new behavior.

>> It is quite common that moving from parsing config settings eagerly by
>> calling repo_config() at startup to parsing them lazily via 'stuct
>> repo_settings' causes regressions like this. We really should find a way
>> to address that before moving more settings into 'struct repo_settings'
> 
> Very true.  If we know the set of things we parse early and have a
> way to say "this command only X, Y, and Z matters (but not W)", then
> the above cat-file example can omit the attributesFile from the "we
> care" set.

That would be nice, but it would be painful to implement for commands 
like cat-file which sometimes need to read core.attributesFile and 
sometimes don't depending on which options they're passed as we 
typically read the config before parsing the command line options. We'd 
also need to be careful to update the list when adding new features 
which depended on config variables that were not previously used by that 
command.

> I think overusing repo_settings is a disease.  Moving a singleton
> global to per repository (by adding to struct repository) is one
> thing and it is very welcome.  But changing the way configuration
> variables are parsed (e.g., what used to be parsed by only those who
> care about is now parsed by everybody, or vice versa) needs to be
> handled carefully.

Indeed

Thanks

Phillip