Superseded by D8694, D8695, D8697 and D8698.

I don't have a good explanation for why we see it, just that it is a coin flip under load for me. Otherwise I would have gone into more details.

I can split the patch in some smaller parts. Is the API change itself as goal reasonable?

I'm still working on this changeset to assert the "extra bytes in node" nonsense documented in some places, if someone has a case to actually trigger that it would be very helpful.

Addressed most of the issues raised, some of them were considered and rejected (re items()).

phabsend didn't properly create a stack here, which is why it doesn't apply by itself.

Please use a tulpe for editopts so that it is a constant by itself.

Further analysis shows that the main issue here is that the topic exchange triggers a rebuild of the "served-topic" branchmap without a real reason. That doesn't mean the change isn't useful in general.

Further analysis shows that the main issue here is that the topic exchange triggers a rebuild of the "served-topic" branchmap without a real reason. That doesn't mean the change isn't useful in general.

No strong reason, it just was easier to start as base line.

This currently triggers a warning about using ui in util, so consider this a RFC for dealing with that.

Yes, that version is better.

At the very least, there is missing a word like "also" or "still", but it doesn't really tell the *why*. That's what would make a comment here useful, IMO. I.e. the important point is that a repo path is a relative path from the repo root, so if the cwd is empty, no additional work is necessary.

I find the comment confusing. Maybe something like Relative paths below the top-level of the repository need more processing.?

While I don't really agree with the design interpretation of why the server should send mandatory, I don't care enough in this case. With the update, bundle and pull/push are getting different flags.

Will fix the real problem.