Re: [PATCH] setup: tweak upgrade policy to grandfather worktreeconfig

[Jonathan Nieder <jrnieder@gmail.com>]

Hi,

Derrick Stolee wrote:
> On 7/13/2020 4:18 PM, Junio C Hamano wrote:

>> The "fail and warn" approach introduced in the previous step may be
>> with smaller impact to the codebase, but
>>
>>  - it requires the end-user to read, understand and execute the
>>    manual upgrade
>>
>>  - it encourages to follow the same procedure blindly, making the
>>    protection even less useful
>>
>> Let's instead keep failing hard without teaching how to bypass the
>> repository protection, but allow upgrading even when only the
>> worktreeconfig extension exists in an old repository, which is
>> likely to be set by a broke version of Git that did not update the
>> repository version when setting the extension.
>
> This is a more subtle way to handle the case. In fact, it
> silently makes extensions.worktreeConfig work as it did in 2.27,
> which means users will not have any troubles after upgrading.

I'd like to propose a different endgame:

Instead of looking at `extensions.*` settings one by one to see how
various implementations handled them with repositoryFormatVersion=0,
what if we treat repositoryFormatVersion=0 as a synonym of version=1?

That is:

1. in new repositories, set repositoryFormatVersion = 1, since (1) Git
   versions in the wild should all already know how to handle it and
   (2) as we've learned, other Git implementations need to understand
   some of extensions.* anyway

2. whenever setting any extensions.*, automatically upgrade to
   repositoryFormatVersion = 1

3. when in an existing repository with extensions.* set and
   repositoryFormatVersion = 0, act as though repositoryFormatVersion = 1

4. document this new behavior with repositoryFormatVersion = 0 in
   Documentation/technical/repository-version.txt

This way, the result is simpler than where we started.

Unfortunately, we are not even that consistent about what to do with
extensions.* settings when repositoryFormatVersion = 0 today.  So
we'll have to exercise some care and analyze them one by one to make
sure this is safe (and if it isn't, at *that* point we'd come up with
a more complex variant on (2) and (3) above).

It's too late to go that far for 2.28.  It would be tempting to try a
simple revert of 14c7fa269e4 (check_repository_format_gently(): refuse
extensions for old repositories, 2020-06-05) to get back to tried and
true behavior but that does not do enough --- it still produces an
error when trying to upgrade repository format when any extensions are
set.  So how about such a revert plus Junio's patch plus the analogous
change to Junio's patch for

  extensions.preciousObjects
  extensions.partialClone

?

Thanks,
Jonathan