[MERGE] new merge-type=lca

Aaron Bentley abentley at panoramicfeedback.com
Sat Dec 29 00:29:21 GMT 2007


Hi all,

This patch implements support for the LCA merge type, which I believe is
a great general-purpose merge algorithm, with excellent criss-cross
support.  I'm very excited about it, so grab a coffee, and I'll yabber on.

In the general case (no criss-cross), it is a three-way merge.  When
there is a criss-cross at the tree level, but not for the particular
file, it is still a three-way merge.  When there's a file-level
criss-cross, it's superior to a three-way merge.

Here's how it works:

First, we compare the files we are trying to merge, and find the lines
that differ.  Next, we try to determine why they differ; this is
essential to the merge operation, because it affects how we resolve the
differences.  In this merger, there are three possible outcomes:

1. The line was added in this version: "new-this"
2. The line was deleted in the other version "killed-other"
3. The line was preserved as part of merge resolution in this version,
but deleted in the other version. "conflicted-this"

Option 3 is new, but I believe it is essential.  When each side has made
a conflicting merge resolution, we should let the user decide how to
combine the two resolutions, i.e. we should emit a conflict.  We cannot
silently drop the line, or silently keep the line, which can happen if
we choose options 1 or 2.  If we choose options 1 or 2, there's also a
possibility that a conflict will be produced, but no guarantee.  We need
a guarantee, which is why we need a new possible outcome.

To decide whether a line is "new-this", "killed-other" or
"conflicted-this", we compare this version against the versions from
each "least common ancestor" (LCA), in graph terminology.  For each LCA
version, if the line is not present in the LCA version, we add it to the
"new" set.  If the line is present in the LCA version, we add it to the
"killed" set.

When we are done going through each LCA version, each unique line will
be in at least one of the sets.  If it is only in the "new" set, it's
handled as "new-this".  If it is only in the "killed" set, it's handled
as "killed-other".  If it is in both sets, it's handled as
"conflicted-this".

The logic here is a bit tricky: first, we know that the line is present
in some, but not all, LCAs.  We can assume that all LCAs were produced
by merges of the same sets of revisions.  That means that in those LCAs,
there were different merge resolutions.  Since THIS and OTHER disagree
about whether the line is present, those differences have propogated
into THIS and OTHER.  Therefore, we should declare that the lines are in
conflict, and let the user handle the issue.

Now, in the common case, there's a single LCA, and LCA merge behaves as
a three-way merge.  Since there's only one LCA, we cannot get the
"conflicted-this" outcome, only "new-this" or "killed-other.  Let's look
at the typical description of three-way merges:

THIS | BASE | OTHER | OUT
-----+------+-------+----
A    | A    | A     | A
-----+------+-------+----
A    | B    | A     | A
-----+------+-------+----
A    | B    | B     | A
-----+------+-------+----
A    | A    | B     | B
-----+------+-------+----
A    | B    | C     | *conflict*

Now, let's assume that BASE is a common ancestor, as is typically the
case.  In fact, for best-case merges, BASE is the sole LCA.

We always pick the version that represents a change from BASE, if there
is one.  For the AAAA line, there is no change, so the output is
rightfully BASE/THIS/OTHER.  For ABAA, the THIS and OTHER are changes
from BASE, and they are the same change so they both win.  (This case is
sometimes called convergence.)  For ABBA, THIS is a change from BASE, so
THIS wins.  For AABB, OTHER is a change from BASE, so OTHER wins.  For
ABC*, THIS and OTHER are both changes to BASE, but they are different
changes, so they can't both win cleanly.  Instead, we have a conflict.

Now in three-way merging, we typically talk about regions of text.  In
weave/knit/newness/lca merge, we also have regions.  Each contiguous
group of "unchanged" lines is a region, and the areas between them are
also regions.

Let's assign a to THIS and b to OTHER.  "unchanged" regions represent
the AAAA or ABAA cases; it doesn't matter which, because the outcome is
the same regardless.  Regions which consist of only "new-a" or
"killed-a" represent the ABBA case.  Regions which consist of only
"new-b" or "killed-b" represent the AABB case.  Regions which have
(new-a or killed-a) AND (new-b or killed-b) are the ABC* case-- both
sides have made changes, and they are different changes, so a conflict
must be emitted.

This is what I mean when I say that it is a three-way merge in the
common case; if there is only one LCA, then it is merely an alternative
implementation of three-way.  (One that happens to automatically do
--reprocess, ftw).

Why a new name:
1. It was time.  Although knit / annotate merge and newness merge have
tried to emulate the behavior of the original weave merge algorithm,
--merge-type=weave hasn't been based on weaves for a long time.
2. Behavior differences.  This algorithm should behave like a three-way
merge in the common case, while its predecessors did not.  It also has
explicit support for handling conflicting merge resolutions, so it
should behave better in criss-cross merge scenarios.

Performance:
Unlike the current weave merge implementation, lca merge does not
perform any whole-history operations.  LCA selection should scale with
the number of uncommon revisions.  Text comparison time should scale
mO(n^2), where m is the number of LCAs, and m is the number of lines in
the file.  The current weave merge compares each uncommon ancestor,
potentially several times, so it is >= kO(n^2), where k is the number of
uncommon ancestors.  So "lca" should beat "weave" both in history
analysis time and in text comparison time.

Possible flaws:
1. Inaccurate LCA selection.  Our current LCA algorithm uses
Graph.heads(), which is known to be flawed.  It may occasionally give
bad results.  This risk is mitigated by the fact that the per-file
graphs tend to be simpler than the revision graph.  And since we're
already using this LCA algorithm, this is not an additional risk.  I
hope that John Meinel will soon have a fixed version of Graph.heads for us.
2. False matches.  Weaves have a concept of line identity, but knits and
later formats do not.  So a line may appear to be common to two files,
when in fact it was introduced separately into each for entirely
different reasons.  This risk is the same for three-way merging.  It is
mitigated by using Patience sequence matching, which a
longest-common-subsequence match.

I think this could be a great merge algorithm, and a candidate to make
our default, but this work would not have been possible without the work
of others, especially:
- Martin's weave merge and knit/annotate merge algorithms.
- Bram Cohen's discussions of merge algorithms
- Andrew Tridgell's dissection of BitKeeper merge
- Nathaniel Smith's analysis of why criss-cross histories necessarily
  produce poor three-way merges.

Aaron
-------------- next part --------------
A non-text attachment was scrubbed...
Name: bzr-mail-EV7G-T.patch
Type: text/x-patch
Size: 27744 bytes
Desc: not available
Url : https://lists.ubuntu.com/archives/bazaar/attachments/20071228/e1abd294/attachment-0001.bin 


More information about the bazaar mailing list