On Wed, Jan 23, 2013 at 9:51 AM, Julian Foad <[email protected]> wrote:
> Ivan Zhakov wrote:
>
>> I was testing recent changes in ra_serf update editor and noticed that
>> reintegrate-like merges for long living branches are extremely slow.
>> Client requests server for diff between branches with respect to
>> ancestry and servers reports no-op txdelta for every file that was
>> changed in original branch. Then for every such file client retrieves
>> text and properties.
>>
>> For example try to reintegrate ev2-export branch back to trunk.
>
> Thanks Ivan, that's very interesting. I'll take a look, since I have just
> been working on that code.
>
> I exposed this as a separate option because the two meanings of the previous
> 'ignore_ancestry' option were conflated, but I don't deeply understand what
> happens when this option is specified. I know what it's supposed to mean at
> a basic level: "diff a pair of unrelated nodes as if they are related". What
> I don't know is how well it's implemented and whether it's really useful when
> merging.
>
>> What is the purpose of diff_ignore_ancestry for merges? Can we default
>> it to FALSE?
>
> I assume one of the purposes is if your work flow has been such that a file
> is sometimes replaced (without copy-from)
Any replacement is a problem, regardless of copy-from. For example:
### Starting with a vanilla Greek tree,
### make a branch:
>svn copy A branch
A branch
>svn ci -m "make branch"
Adding branch
Committed revision 2.
### Now replace two files on the branch,
### one with history, one without:
>svn del branch\B\lambda
D branch\B\lambda
>svn move branch\D\gamma branch\B\lambda
A branch\B\lambda
D branch\D\gamma
>svn del branch\mu
D branch\mu
>echo replacement without history > branch\mu
>svn add branch\mu
A branch\mu
>svn st
R + branch\B\lambda
> moved from branch\D\gamma
D branch\D\gamma
> moved to branch\B\lambda
R branch\mu
>svn ci -m "Replace some files on the branch"
Replacing branch\B\lambda
Deleting branch\D\gamma
Replacing branch\mu
Transmitting file data .
Committed revision 3.
>svn up -q
### Sync merge from 'A' to 'branch':
>svn merge ^^/A branch
--- Recording mergeinfo for merge of r2 through r3 into 'branch':
U branch
>svn ci -m "Synch the branch with its parent"
Sending branch
Committed revision 4.
### Now do a "reintegrate" style auto merge
### without Ivan's patch. The replacements
### are merged as expected:
>svn merge ^^/branch A
--- Merging differences between repository URLs into 'A':
R A\B\lambda
R A\mu
D A\D\gamma
--- Recording mergeinfo for merge between repository URLs into 'A':
U A
>svn st
M A
R + A\B\lambda
D A\D\gamma
R + A\mu
### Now try the same merge with Ivan's patch.
### History is lost as the merge treats the replacements
### as simple modifications:
>svn revert -Rq .
>svn merge ^^/branch A
--- Merging differences between repository URLs into 'A':
U A\B\lambda
U A\mu
D A\D\gamma
--- Recording mergeinfo for merge between repository URLs into 'A':
U A
>svn st
M A
M A\B\lambda
D A\D\gamma
M A\mu
>, and sometimes a new file added on one branch has been added on the other
>branch by a simple add (without copy-from). In that sort of work flow, the
>'diff_ignore_ancestry' would cause Subversion to do diffs between different
>versions of a file, and that might well help in merging the changes. Without
>the 'diff_ignore_ancestry' option, Subversion would treat the file as
>'replaced', and so you'd be likely to get a tree conflict when you try to
>merge it.
> Another case that demonstrates the option better (but not in an automatic
> merge) would be if you have two tree snapshots (vendor drops of product
> version 1.1 and 1.2, say) and you want to merge the difference between them
> into your working copy.
>
> Can we default it to ... I think you mean TRUE, in an automatic merge?
> Perhaps. But let me take a look at the code to see if this is buggy
> behaviour, first. Also I'll try to think whether it makes sense to default
> to TRUE and whether we'd then want a user interface override like the
> existing 'svn diff --notice-ancestry' option.
>
>
>> I've made simple patch that changes unconditional diff_ignore_ancestry
>> = FALSE to TRUE in merge_cousins_and_supplement_mergeinfo() call and
>> everything seems to be working and ALL tests are passing.
>
> You mean in the call to merge_cousins_and_supplement_mergeinfo() in
> do_automatic_merge_locked(), I assume.
>
> - Julian
--
Paul T. Burba
CollabNet, Inc. -- www.collab.net -- Enterprise Cloud Development
Skype: ptburba
