Junio C Hamano wrote:
> Victoria Dye <vdye@xxxxxxxxxx> writes:
>
>> I think the decision of whether to go with this approach or the alternative
>> listed below depends on the validity of a 'repository' without a gitdir.
>>
>> As far as I can tell, there is an implicit conflict between the changes in:
>>
>> 1. b66d84756f (commit-graph: respect 'commitGraph.readChangedPaths',
>> 2020-09-09)
>> 2. 44c7e62e51 (repo-settings: prepare_repo_settings only in git repos,
>> 2021-12-06) (as you pointed out in your message)
>>
>> The former says that commit-graph should use a repository setting (implying
>> it needs a valid repository), and the latter says that you need a valid
>> gitdir to get repository settings.
>>
>> So to me, how to proceed depends on whether a repository can be "valid"
>> without a gitdir or not:
>
> Sorry, I do not get it. What does "a repository without a git dir"
> look like? It does not make any sense to me. A repository without
> working tree, I can understand.
I think that answers my question - if it doesn't make sense to have a
"repository without a git dir", then the code shouldn't allow
'the_repository' to be used without a valid 'the_repository.gitdir'. The
'BUG()' in 'prepare_repo_settings()' enforces that condition right now, so
removing it like this RFC does would make the tests fragile and the code
more prone to future bugs.
That said, if it's conceptually sensible for 'fuzz-commit-graph' to work
without a repository, then it could be updated to get its defaults from
somewhere other than 'the_repository' when the repo doesn't exist (like what
I mentioned in [1] ("If the 'repository' really is..."), or Taylor's
suggestion in [2]).
[1] https://lore.kernel.org/git/fcfdcbb9-761a-0d34-7d36-61e0ef279922@xxxxxxxxxx/
[2] https://lore.kernel.org/git/Yjt6mLIfw0V3aVTO@nand.local/
