Re: What can Debian do to provide complex applications to its users?
Here is a different straw man, which I think might be similarly effective
and a lot less work:
On Tue, 27 Feb 2018 at 14:13:41 +0100, Didier 'OdyX' Raboud wrote:
> As Debian, we
> are insisting that our releases ideally only contain a single version of a
> software, that we insist is made available at system-level.
> In other words, vendorization is the tool that allows developers to get rid of
> distribution constraints and get on with their development through installing
> the dependencies from their ecosystem as they see fit (non-root), in the
> (eventually precise) version they need.
I can't help wondering whether vendorizing dependencies (embedded code
copies) would be a better route for ecosystems like npm that we might
categorise as "aggressively modular" - package the app, but not the
dependencies, and treat the vendorized/embedded dependencies as part
of the app. Not embedding or duplicating libraries is already an ideal
that we have to compromise on for pragmatic reasons - browsers use
system libraries in testing/unstable but gradually make increasing use
of embedded code copies in stable, drifting further away from the
"no embedded code copies" ideal as time goes on and the latest browser
version becomes increasingly unbuildable against the increasingly old
libraries in stable. It's also how many Rust dependencies are already
managed, as far as I'm aware.
Similarly, we are willing to tolerate embedded code copies for C libraries
that are specifically designed to be "copylibs", such as GNU's gnulib,
GNOME's libglnx, libgd and (historically) libegg, and the stb_image
micro-library used in multiple game engines, as well as many Autoconf
macros, CMake modules and other supporting files. We do that because these
copylibs are explicitly unstable (updating may require code changes in
the consuming code), or because consumers require a very new version,
or both. Is that really a million miles away from the npm ecosystem?
The obvious retort is "but what about security updates?". Well, what
about security updates? What would happen if there was a security
vulnerability in gnulib code? I'm fairly sure the answer is that the
upstream maintainers of packages that had imported the relevant code,
or the downstream maintainers of those packages, would be expected to
patch the vulnerability - just like what would happen if they had
pasted individual functions into their codebase. (Some nodejs modules
*are* individual functions, for that matter.)
Also, the security team specifically don't provide security
support for libv8, which apparently extends to node-* packages like
<https://security-tracker.debian.org/tracker/CVE-2015-8855>, so it's
hard to see how tolerating embedded code copies of nodejs modules in
particular would make their security support situation a whole lot worse:
it's already the case that the upstream and downstream maintainers of
these modules (or the applications that bundle them, or both) provide
the only security maintenance they'll get. In practice, this isn't as
awful as it first appears, because nodejs modules are often very small,
so an individual nodejs module is relatively unlikely to contain security
vulnerabilities even if its defect density is high, simply because there
isn't very much code to be vulnerable.
There is nothing to stop us from enforcing our quality and freeness
standards equally thoroughly for embedded/vendorized code copies -
they're just code, after all. If the process works for the main package,
then extending it to a bundle of vendorized dependencies that share
an orig tarball (or cluster of multiple orig tarballs in 3.0 source
formats) is a matter of scale rather than an entirely separate process,
and if a maintainer or maintainer team can't cope with that scale,
then the same actions need to be taken as for any other package that
can't be maintained to our community standards.
I'm not saying that we should vendorize everything: for large codebases
that are an identifiably separate module with an API (like libjpeg
and zlib) and/or have a history of security vulnerabilities that can
be fixed centrally without breaking ABI (like libjpeg and zlib), it is
of course often absolutely the right course of action to use a system
copy. Similarly, if a maintainer feels that "unvendorizing" a particular
part of the code and using a shared system-wide copy would make their
life easier, they should be welcome to do so.
I'm just not sure that taking the rules we follow for C/C++ shared
libraries and applying them to every other language's ecosystem is either
feasible or desirable - nodejs libraries are not the same as C libraries,
and their tradeoffs are not the same tradeoffs.
This approach is also not a million miles away from the approach taken
in Flatpak, where every dependency is either the runtime vendor's
responsibility (system libraries like libjpeg, zlib, SDL, GTK+) or the
app vendor's responsibility (embedded/bundled/vendorized libraries),
and if the app vendor has chosen poorly, it's their responsibility to fix
the resulting CVEs. (See the videos of my Debconf 17 or FOSDEM 2018 talks
for more background on Flatpak.) I don't think this is a terrible model.
Firefox 58 uses system zlib, bz2, libffi, libevent everywhere, but
uses embedded code copies of hunspell, nspr, nss, sqlite in stretch
and older, and an embedded code copy of vpx in jessie and older.
 https://git.gnome.org/browse/libgd, not to be confused with the
graphics library of the same name