From: "Junio C Hamano" <gitster@xxxxxxxxx>
Sent: Thursday, April 18, 2013 1:13 AM
"Philip Oakley" <philipoakley@xxxxxxx> writes:
How about
* E.g. git gui uses the extended regular expression "^git version
[1-9]+(\.[0-9]+)+.*"
* to check for an acceptable version string.
The ERE is from git-gui.sh:L958.
That is exactly the kind of guarantee we do _not_ want to give.
So you are trying to avoid giving _any_ "guarantee" about what visible
manifestation a user may obtain about a system that passes the test
suite we have set out. I was hoping we could give a basic minimum
indication of the expected style of the version string for a git which
*passes* our tests. But like you say, it doesn't form a real guarantee -
caveat emptor still applies.
... Hence my suggestion of the basic test that a "passing" git
would produce a consistent version string.
I have been assuming that you are trying to avoid an exchange like
this one we had in the past:
http://thread.gmane.org/gmane.comp.version-control.git/216923/focus=217007
In a sense yes. As David noted, others do attempt to trust us via our
current version string format. I thought it worthwhile (given my earlier
'mistake' in 216923/focus=216925) to suggest such a basic indication of
our current string style.
I also have been assuming that you are pushing to limit the possible
versioning scheme, but I do not know what that extra limitation
would achieve in the real world.
By sticking to the pattern "git gui" happens to use, "git gui" may
be able to guess that the next version of Git says it is verison
"1.8.3". That is the _only_ thing your "test" buys.
But having parsed the "1.8.3" substring out of it, "git gui" does
not know anything about what 1.8.3 gives it. As far as it is
concerned, it is a version whose "git version" output it has never
seen (unless it has been kept up to date with the development of
Git).
You are focusssing on the wrong side of issue (from my viewpoint).
If my earlier patch had gone in, it would have passsed our tests but not
played nicely with the community tools. That would have been IMHO a
regression, so from my viewpoint I believed it was worth having a test
to avoid such a 'regression'.
By matching against "git version [1-9]+(\.[0-9]+)+", it is accepting
that future versions may break assumptions it makes for some known
versions (which is warranted) and all future versions (which is
unwarranted) of Git. Maybe the version 2.0 of Git adds all changes
in the directory "d", including removals, when you say "git add d",
but it may have assumed that with Git version 1.5.0 or newer, saying
"git add d" would result in added and modified inside that directory
getting updated in the index, but paths removed from the working
tree will stay in the index.
If it was a gross incompatibility then (a) we are likley to be
signalling it for a while, and (b) other tools would need updating as
well, and they would hope that they could 'read' the version in a
consistent style. We would also still have the choice of changing our
existing string style, which would explicitly signal a change via the
test fail.
The only thing the scripts that read from the output of "git
version" can reliably tell is if it is interacting with a version of
Git it knows about. If it made any unwarranted assumption on the
versions it hasn't seen, it has to be fixed in "git gui" _anyway_.
Of course, we would not change the output of "git version"
willy-nilly without good reason, but that is a different topic.
Ah. I thought it was the [original] topic. I was calibrating the
willy-nillyness ;-)
So I do not know what you want to achieve in the real world with
that extra limitation on the "git version" output format.
Maybe you are proposing something else. I dunno.
I was just using a slightly different philosophical approach.
--
Philip
--
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html