****************
Joining branches
****************

(I think this is pretty brilliant. :-)

Branches diverge when people create more than one changeset following
on from a common ancestor::

  A:   0 ------- 1
  B:    \------- 2

We also allow branches to reunite.  This means that all the decisions
taken on multiple branches have been reconciled and unified into a
single successor::

  A: 0 ------- 1 ----- 3
      \               /
  B:   \------ 2 ----/

The predecessor of 3 is 1, in the sense that it was created on that
branch.  We could have created the exact same state as a successor to
2, and we can move that state onto branch 2 without any loss of
information.

(One thing we can do here is just delete B.  Because all of the work
there has been merged onto A, this will not lose anything.  We might
do this if the purpose of B has been achieved, such as completing a
feature or bug.  But if the work is still in progress, we might keep
it around.  It makes little difference whether we decide to do new
work in a branch called B or make a new one called C.)

That is to say that 3 can be perfectly (trivially) merged onto B,
with, say ``bzr push``, ``bzr pull`` or ``bzr update`` (whatever name
works best).  Perfectly merged means that we know there will be no
conflicts or need for manual intervention, and that we can just
directly store it without forming a roll-up changeset.

I think we might also like the choice of merging A onto B, rather than
pulling the changeset.  That causes a new changeset to be created on
B, noted as a successor of 2 and 3::

  A: 0 ------- 1 ----- 3 ------+
      \               /         \
  B:   \------ 2 ----+---------- 4

One complication is that 3 is probably stored in A's history as a
patch relative to 1; we can't just move this representation across.
Instead, we need to recalculate the delta from 2 to 3 and store that.

Despite that the delta is stored differently, the original signature
on 3 should still be valid.  So it must be a signature of the tree
state, not the diff.

Note from `Kernel Traffic discussion`__:

__ http://www.kerneltraffic.org/kernel-traffic/kt20030323_210.txt

    But anyway, what made Bitkeeper suck less is the real DAG
    structure.  Neither arch (http://arch.fifthvision.net/bin/view)
    nor subversion seem to have understood that and, as a result,
    don't and won't provide the same level of semantics. Zero hope for
    Linus to use them, ever. They're needed for any decently
    distributed development process.

This in turn suggests that possibly deltas should be stored separately
from the commits that create them.  Commits name points of
development; deltas describe how to get from one to the other.

The separation is nice in allowing us to send just a delta when
diffing trees or recording for undo.  We might want to compute many
deltas between different trees.

Is this a problem?  Does this ignore Tom's advice about the primacy of
storing changesets?

Splitting them is probably good, but then what manifest is stored in
changesets?  We don't want to store the manifest of the whole tree if
we can avoid it.  So I suppose the changeset just gives the hash of
the manifest, and the manifest then can be stored separately, possibly
delta-encoded.