Note: that's an unrelated change.

Agreed. As we talked earlier today, this will need a change to swh.storage.get_storage (D689).

It's the contents, at "version 1" of the DB (version 0 = before the loader runs, version 1 = after it runs for the first time, version 2 = after it runs for the second time).

CONTENTS_v{0,1,2} etc... would have been more appropriate...
Let's keep this on the side for now.

Do you need to?
That's the part reading the disk for configuration file.

So, that could create divergence on dev box if they have some configuration files in the right places ({/etc/softwareheritage/,~/.swh,~/.config/swh}/loader/git-{loader,updater}.yml))

I need a way to obtain content_packet_size and its many friends.

In other modules, we duplicate (for now) that information in the TestLoader (TestGitLoader for example).

Side-effects/configuration duplication trade-offs.

Instead of duplicating base configuration, I would rather use this implementation (see D768)

def parse_config_file(self, *args, **kwargs):
        config = super().parse_config_file(*args, **kwargs)
        config['storage'] = {'cls': 'memory', 'args': {}}
        return config

I'm the source of disagreement here.
I'd like to avoid the call to the original parse_config_file method here to reduce the disk reading part.
(I've said as much in D768 indeed).

If you take a look at the discussion just below this remark, you'll see the detail ;)