[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Feedback on 3.0 source format problems

Guillem Jover writes ("Feedback on 3.0 source format problems"):
> I'm interested in what things people still find so off-putting to the
> point of not wanting to use the new 3.0 source formats, or what makes
> people use them in anger and similar (if people would state which one
> of these apply that would be helpful). All these including objective
> and subjective issue. And even existing bug references would be fine.

Thanks for explicitly inviting critical opinions!

AIUI the only significant 3.0 format supported in Debian right now is
`3.0 (quilt)'.

I think it is very poor.  Its only virtue is that it is better than
in-tree ad-hoc patch systems implemented in debian/rules, which it has
(thankfully!) almost entirely replaced.

IMO it's not a `source package format'.  This is because a `source
package format' would be a way to represent and transport a package's
source code tree.  But `3.0 (quilt)' puts extraneous data in the
source tree (in debian/patches/), and requires those files to be kept
in step.  Or to look at it another way, it insists on execution of
special commands after editing files.

Looked at another way, it is trying to be a version control system,
layered on top of the Debian archive.  But it is only about a quarter
of a VCS.  There are no formal interfaces to do proper VCS operations.
If there is a formal interface, it is quilt(1) (which is itself very
poor.  NB that is not quilt's fault: quilt inevitably has a hard job
because can't make enough assumptions).

I think if we want to be storing patch queues we should be doing so in
a real version control system.  Indeed, most people are doing that.
For now many people are using `3.0 (quilt)' as an interchange format,
but ideally we would switch to a useable git workflow that did a
similar thing, and then we could use git as our history interchange

Additionally, `3.0 (quilt)' unfairly privileges the Debian Project: it
provides facilities (which apparently some people find helpful) for
maintaining a delta from an upstream tarball.  But because it excludes
patches to debian/* from its patch series, it cannot be used the same
way by a Debian derivative to maintain their packaging patch stack, on
top of their own upstream (which might be Debian or might be an
intermediate derivative).

There are other 3.0 formats, which are all quite different.  The only
interesting one of these which is a serious contender is `3.0 (git)'.
IMO `3.0 (git)' is a very silly way of transporting git branches
about.  It doesn't solve any underlying technical problems; it just
tries to wedge git into the ftp.debian.org-shaped hole.

`3.0 (native)' exists too but I don't quite understand the point of
that.  1.0 could be extended to support other compression formats, and
it does not make any logical sense that ignore rules depend on the
source format.

I have already said that I would like a source format which can
faithfully represent every possible reasonable tree (let us say, every
git tree).

We need such a format to retain the space and bandwidth advantages of
.orig tarballs.  It needs to be very stable and not to accidentally
grow new incompatible features (as `3.0 (quilt)' has done due to its
reliance on patch).

If we manage to store our the revision control history elsewhere, in a
proper revision control server, we do not need the archive to contain
the revision control history.

I think my proposal for `3.0 (rsync)' would meet these requirements
and I'm still sad it won't make stretch.  It means we can't use that
for developing buster.


Ian Jackson <ijackson@chiark.greenend.org.uk>   These opinions are my own.

If I emailed you from an address @fyvzl.net or @evade.org.uk, that is
a private address which bypasses my fierce spamfilter.

Reply to: