[svn.haxx.se] · SVN Dev · SVN Users · SVN Org · TSVN Dev · TSVN Users · Subclipse Dev · Subclipse Users · this month's index

Re: RFC: change 'svn merge' default behavior.

From: Sander Striker <striker_at_apache.org>
Date: 2004-01-15 18:42:47 CET

On Thu, 2004-01-15 at 18:25, Ben Collins-Sussman wrote:
> Gentlefolk,
>
> Karl and I want to propose a very simple change to 'svn merge' default
> behavior in svn 0.37. Allow me to explain.
>
>
> 1. We have a problem
>
> Early this week I merged all of the book changes from trunk to the 1.0
> branch. The merge seemed simple enough -- a bunch of files got
> patched, but one had a conflict. I couldn't figure out why there'd be
> a conflict... the book never changed on the branch. After some
> investigation, I discovered that totally bogus hunks of text had been
> applied to the wrong files! Most the local mods from 'svn merge' were
> bogus.

Ouch!

> What happened? My trunk changes involved renames; I moved ch06->ch07,
> ch07->ch08, and ch08->ch09. 'svn merge', just like 'svn diff',
> ignores file ancestry (copy history) by default. So it was applying
> ch07 changes from trunk to a similarly named (but totally unrelated!)
> file on the branch. After reverting the bogus merge, I ran 'svn merge
> --notice-ancestry' and got the correct set of changes: a bunch of
> deletes and adds of identically-name files. Everything was fine.
>
>
> 2. The history of this issue
>
> Once upon a time, diff and merge used to both notice ancestry by
> default. But this angered users of 'svn diff'.

I still wonder how many people we are talking about here.

> When running 'svn
> diff URL1 URL2', people would be annoyed that diff was being "too
> smart". It would notice that two identical file paths were unrelated,
> and show a delete of one file (all minuses) and the add of the other
> file (all pluses). But when running 'svn diff', the user usually
> doesn't give a darn about whether identically-named files are related
> or not: they just want to see the diff between them! (The diff format
> doesn't understand tree changes anyway,

yet..

> so why include that
> information?)
>
> So, we did the right thing: we made 'svn diff URL1 URL2' ignore
> ancestry by default, and "dumbly" diff identical file paths.

I still am not supporting this decission. svn diff TARGET1 TARGET2
should indeed ignore history. However, svn diff -r X:Y TARGET should
notice history. Hmpf, we need that peg revision proposal implemented.

> But in
> the interest of consistency, folks decided that diff and merge should
> behave the same. So both commands now ignore ancestry by default, and
> require a long option to notice ancestry.

Merging should always notice ancestry by default. The only time it
is interesting to not notice ancestry is vendor branch merging. This
should be uncommon enough to warrant --ignore-history.

>
> 3. The proposal
>
> Very simple: allow 'svn diff' continue to ignore ancestry by default,
> but change 'svn merge' to notice ancestry by default.

Does this mean that svn diff -r X TARGET compares the WC against the
URL@X instead of the file in the WC at version X?

> (The patch for this involves no API changes; internally, the diff and
> merge client functions already have the ability to ignore ancestry or
> not. The patch is a trivial change to cmdline client's default
> behaviors only. A long option can always be used to override either
> subcommand's default behavior.)
>
> Yes, I understand that this breaks our nice consistency between diff
> and merge. But Karl and I think the risk for producing munged local
> mods here is just too high.

+1 on that assessment.

> 99% of the time, when people are merging changesets from one branch to
> another, people want ancestry to be noticed and copies to be honored.
> It's dangerous and foolish to default to the 1% use-case. Keep in
> mind, I wrote the *book* chapter about branching and merging, and I
> still got burned by merge's default behavior. It's not always so
> obvious to detect the munged local mods... if I hadn't received a
> conflict, I might have accidentally committed garbage.

Which says something about our merge code. It sometimes is a bit too
smart (in not taking context into account). The number of lines of
context to take into account when deciding something is a conflict
or not should be configurable. And probably default to 3 as CVS does.
Anyway, that's a different topic.

> So yes, it's "nice" to have consistent default behaviors in diff and
> merge, but this route leads to either (1) 'svn diff' not showing diffs
> by default (very user-unfriendly), or alternately (2) a significant
> risk of 'svn merge' creating bogus local mods by default. In my
> mind, neither evil is worth the "niceness" of consistency. It's hard
> to enumerate the concrete benefits of consistency, except,
> well... "it's consistent."

The alternative is... a config option for both diff and merge,
defaulting to ignore-ancestry for diff and notice-ancestry for merge.
People can then set the default to their preference. Although I must
admit I don't really want to add yet another config option.

Sander

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@subversion.tigris.org
For additional commands, e-mail: dev-help@subversion.tigris.org
Received on Thu Jan 15 18:43:29 2004

This is an archived mail posted to the Subversion Dev mailing list.