Re: [PATCH] difftool: add support for an extended revision syntax

From: Junio C Hamano <gitster@pobox.com>
Date: Mar 23, 2009, 23:22 UTC
Message-ID: <7vhc1j6czf.fsf@gitster.siamese.dyndns.org>
In-Reply-To: <20090323163343.GB16258@gmail.com>

David Aguilar <davvid@gmail.com> writes:

> Keep in mind that the syntax that this patch added does not have file~3
> = HEAD~3.  file~3 means finding file as it existed 3 changes-to-file
> ago, which is != to HEAD~3 if file did not change in the last 3 commits.

If your motive is to introduce inconsistency to the UI by adding this kind of new notation _only to difftool_, I have to reconsider moving it out of contrib/ area.

While I do not fundamentally oppose to add convenient notations for useful concepts, you need to start at making sure if this "three changes ago" is a well defined concept to begin with.

And it is not a well defined concept in a merge-heavy environment, unless you define what you mean by "three changes ago".

If you consider this history:

 ---Y---o---X---M---o mainline = HEAD
               /
   ---A---B---C topic

where A, B, C and X, Y are the only commits that touched the file you are interested in, how do you define 3-changes-ago?

Maybe X was just a totally uninteresting typofix to a comment, while A, B and C were adding a very interesting new feature. Don't forget that M also changes the file from either of its parents (X or C). Does M count as the last change? Or does it not count because it is just a mechanical unconflicting merge? Which one of X or C is the penultimate change? The one with an earlier committer timestamp? Tiebreaking with timestamps is known to be flawed in the presense of clock skew.

For the consistency of the UI, "starting at HEAD, following first-parent ancestry, find N-th commit that touches the path, ignoring all the side branches" MUST be the semantics of a notation that uses tilde followed by number (so file~3 must mean Y in the above picture), because HEAD~3 is defined as "three parents ago, only following the first parent ancestry". Anything else will invite user confusion.

But I do not think it is necessarily useful to follow only the first parent ancestry to find "three-changes ago" (if such a concept exists). If you want a notation that means something else, such as X (because chronologically the commits that touched the file are M, C and X in the ideal world that everybody has well synchronized clock), you shouldn't use tilde-number notation but use something else.