[MERGE] bzr+http:// should always try POSTing to the same location, not lots of child locations.

[Andrew Bennetts]

Aaron Bentley wrote:
[...]
> I think dumb transport operations should be consistent.
> 
> We've defined that get_transport('x').clone('y') is equivalent to
> get_transport('x/y').  This breaks that.  It penalizes behaviour that we
> should be rewarding, because if we can use an existing transport to
> acquire another BzrDir, that is a performance win.
> 
> This is addressing the issue at the wrong layer.  Instead of changing
> the way dumb transports work, implement a SmartBzrDir.  We want to do
> that anyway, don't we?

Perhaps that's slightly cleaner, but there's no SmartBzrDir in bzr.dev yet.  I'm
unconvinced that it would make any practical difference, because it either way
it only affects smart transports.  Anyone relying on edge cases of file-level
operations over the smart transport is on thin ice anyway, because the plan is
that we'll remove those operations.

I'm not sure what you're referring to with "if we can use an existing transport
to acquire another BzrDir, that is a performance win" — what cases do we do that
now that my patch would break?  I still take the same measures as the plain old
HTTP transports to allow multiple transports to reuse the same HTTP connection.

-Andrew.