svn tree conflicts: what am i doing wrong?

merge, svn, tree-conflict

Solution

The problem you've encountered is not due to the 'local copy'[1] per se. The problem lies in the fact that, in revision 3, you copy a mixed-revision working copy (http://svnbook.red-bean.com/en/1.7/svn-book.html#svn.basic.in-action.mixedrevs).

If we run your script right up until you copy 'trunk' to 'branches/a' we find that we have a mixed-revision working copy:

>svn st -v
                 0        0  ?           .
                 1        1 pburba       branches
                 1        1 pburba       trunk
                 2        2 pburba       trunk\colors

So when you copy 'trunk' to 'branches/a', you are actually copying trunk@1 and trunk/colors@2:

>svn copy trunk branches\a
A         branches\a

>svn st -v
                 0        0  ?           .
                 1        1 pburba       branches
A  +             -        1 pburba       branches\a
A  +             -        2 pburba       branches\a\colors
                 1        1 pburba       trunk
                 2        2 pburba       trunk\colors

>svn ci -m "Copy mixed-rev trunk"
Adding         branches\a
Adding         branches\a\colors

Committed revision 3.

We can see this most clearly when looking at the verbose log for r3:

>svn log -v -r3
------------------------------------------------------------------------
r3 | pburba | 2013-03-11 15:31:23 -0400 (Mon, 11 Mar 2013) | 1 line
Changed paths:
   A /branches/a (from /trunk:1)
   A /branches/a/colors (from /trunk/colors:2)

Copy mixed-rev trunk
------------------------------------------------------------------------

Skipping ahead to the problematic merge, we have a "clean" working copy with no mixed-revisions and no local modifications. So far so good:

>svn st -v
                 5        5 pburba       .
                 5        5 pburba       branches
                 5        5 pburba       branches\a
                 5        5 pburba       branches\a\colors
                 5        4 pburba       trunk
                 5        4 pburba       trunk\colors

But if we use the svn mergeinfo command to preview what revisions will be merged we notice that revsion 2 is eligible:

>svn mergeinfo --show-revs eligible ^/trunk branches\a
r2
r4

But wait, revison 2 is the addition of 'colors',

>svn log -v -r2
------------------------------------------------------------------------
r2 | pburba | 2013-03-11 15:43:52 -0400 (Mon, 11 Mar 2013) | 1 line
Changed paths:
   A /trunk/colors

created trunk/colors with red inside
------------------------------------------------------------------------

and we already copied that when we created the branch in revision 3! So why is Subversion trying to merge it again? The reason comes back to the mixed-revision WC-to-WC copy we made. The merge target 'branch/a' knows that it was copied from trunk@1, prior to the addition of 'trunk/colors'. So the merge thinks that revision 2 hasn't been been merged to branch/a and tries to merge this change, adding 'colors' into a 'a' where a file of the same name already exists, which results in a tree conflict:

>svn merge ^/trunk branches\a
--- Merging r2 through r5 into 'branches\a':
   C branches\a\colors
--- Recording mergeinfo for merge of r2 through r5 into 'branches\a':
 U   branches\a
Summary of conflicts:
  Tree conflicts: 1

>svn st
 M      branches\a
      C branches\a\colors
      >   local file obstruction, incoming file add upon merge
Summary of conflicts:
  Tree conflicts: 1

So we tricked Subversion with our mixed-revision WC-to-WC copy, which effectively brought revision 2 to the branch during the copy. So why can't Subversion detect this and skip revision 2? In this case we conceivably could do that, but what if revision 2 included other changes to 'trunk'? e.g.:

>svn log -v -r2
------------------------------------------------------------------------
r2 | pburba | 2013-03-11 15:43:52 -0400 (Mon, 11 Mar 2013) | 1 line
Changed paths:
   A /trunk/colors
   A /trunk/README
   M /trunk

created trunk/colors with red inside, add a README file, and set the
svn:ignore property on trunk
------------------------------------------------------------------------

Subversion doesn't support merging part of a revision to a given path, it's either all-or-nothing.

~~~~~

So how to avoid this problem? As you've already discovered, doing a URL-to-URL copy solves it. Why? Because when the copy source is a URL, it is, by definition, at some uniform revision:

>svn log -v -r3
------------------------------------------------------------------------
r3 | pburba | 2013-03-11 16:02:59 -0400 (Mon, 11 Mar 2013) | 1 line
Changed paths:
   A /branches/a (from /trunk:2)

Create branch 'a' from 'trunk' with a URL-to-URL copy.
------------------------------------------------------------------------

However you can also accomplish the same thing with a URL-to-WC copy:

 svn commit trunk -m 'created trunk/colors with red inside'
-svn copy trunk branches/a
+svn copy ^/trunk branches/a
 svn commit branches/a -m 'created branches/a'

Or simply by updating the WC prior to the WC-to-WC copy in your original script:

 svn commit trunk -m 'created trunk/colors with red inside'
+svn update
 svn copy trunk branches/a
 svn commit branches/a -m 'created branches/a'

Your original solution, the URL-to-URL copy followed by an update, is the best option IMO. The update & WC-to-WC copy and the URL-to-WC copy both allow for the possibility that additional changes can be made to the copy before it is committed -- it's best that the revision in which a branch is created has no other changes besides the copy. That said, all of these option will work fine[2].

[1] In Subversion-speak we usually refer to this as a working-copy-to-working-copy copy, or WC-to-WC copy for short. Contrast this with URL-to-URL, URL-to-WC, or WC-to-URL copies. WC-to-URL copies are also vulnerable to the problem described above. See also 'svn copy --help'.

[2] Of course even using one of these three options still results in a text conflict, but that is expected since we made incompatible changes to the 'colors' file on both trunk and the branch after the branch was created.

>svn merge ^/trunk branches\a
--- Merging r3 through r5 into 'branches\a':
C    branches\a\colors
--- Recording mergeinfo for merge of r3 through r5 into 'branches\a':
 U   branches\a
Summary of conflicts:
  Text conflicts: 1
Conflict discovered in file 'branches\a\colors'.
Select: (p) postpone, (df) diff-full, (e) edit, (m) merge,
        (mc) mine-conflict, (tc) theirs-conflict, (s) show all options: p

>svn st
 M      branches\a
C       branches\a\colors
?       branches\a\colors.merge-left.r2
?       branches\a\colors.merge-right.r5
?       branches\a\colors.working
Summary of conflicts:
  Text conflicts: 1

Problem

Anytime I try to merge in SVN I get piles of tree conflicts. Well, in the case of this sample script just one, but still. ``` #!/bin/bash svnadmin create repo svn checkout file://`pwd`/repo wc cd wc mkdir trunk branches svn add trunk branches svn commit -m 'created trunk and branches' echo red > trunk/colors svn add trunk/colors svn commit trunk -m 'created trunk/colors with red inside' svn copy trunk branches/a svn commit branches/a -m 'created branches/a' echo green >> trunk/colors svn commit trunk -m 'added green to trunk/colors' echo blue >> branches/a/colors svn commit branches/a -m 'added blue to branches/a/colors' svn update svn merge ^/trunk branches/a ``` My result is: ``` Checked out revision 0. A trunk A branches Adding branches Adding trunk Committed revision 1. A trunk/colors Adding trunk/colors Transmitting file data . Committed revision 2. A branches/a Adding branches/a Adding branches/a/colors Committed revision 3. Sending trunk/colors Transmitting file data . Committed revision 4. Sending branches/a/colors Transmitting file data . Committed revision 5. Updating '.': At revision 5. --- Merging r2 through r5 into 'branches/a': C branches/a/colors --- Recording mergeinfo for merge of r2 through r5 into 'branches/a': U branches/a Summary of conflicts: Tree conflicts: 1 ``` I know that SVN isn't known for being merge friendly, but I have to assume that in this case it is my fault somehow. Thanks for any pointers.

Original source