Re: RFC Failover url for fetches?

["Jakub Narębski" <jnareb@gmail.com>]

W dniu 21.10.2016 o 21:03, Junio C Hamano pisze:
> Stefan Beller <sbeller@google.com> writes:
> 
>> So when pushing it is possible to have multiple urls
>> (remote.<name>.url) configured.
>>
>> When fetching only the first configured url is considered.
>> Would it make sense to allow multiple urls and
>> try them one by one until one works?
> 
> I do not think the two are related.  Pushing to multiple is not "I
> want to update at least one of them" in the first place.

Push is/should be 'update all', fetch is/should be 'fetch any'.
I thought that multiple remote.<name>.url values provide this
fallback for fetch, but it looks like it isn't so...

> 
> As to fetching from two or more places as "fallback", I am
> moderately negative to add it as a dumb feature that does nothing
> more than "My fetch from A failed, so let's blindly try it from B".
> I'd prefer to keep the "My fetch from A is failing" knowledge near
> the surface of end user's consciousness as a mechanism to pressure A
> to fix it--that way everybody who is fetching from A benefits.
> After all, doing "git remote add B" once (you'd need to tell the URL
> for B anyway to Git) and issuing "git fetch B" after seeing your
> regular "git fetch" fails once in a blue moon is not all that
> cumbersome, I would think.

One would need to configure fallback B remote to use the same
remote-branch namespace as remote A, if it is to be used as fallback,
I would think.

> 
> Some people _may_ have objection based on A and B going out of sync,
> especially B may fall behind even yourself and cause non-ff errors,
> but I personally am not worried about that, because when somebody
> configures B as a fallback for A, there is an expectation that B is
> kept reasonably up to date.  It would be a problem if some refs are
> expected to be constantly rewound at A (e.g. 'pu' in my tree) and
> configured to always force-fetch, though.  A stale B would silently
> set such a branch in your repository back without failing.

Nb. there is also http-alternates mechanism... which nowadays doesn't
matter anyway, I would think.

-- 
Jakub Narębski