On Thu, Feb 12, 2015 at 01:32:04PM -0500, Stephen Gallagher wrote:
(Logistical note: please keep all replies to this thread on
devel(a)lists.fedoraproject.org)
tl;dr Shall we consider requiring a lesser package review for packages
that are not present on Product or Spin install media?
Despite the title, and the
tl;dr summary, what you are proposing boils
down to a blanket exception on bundling. This is both not enough and
dangerous. I agree that we could relax the rules in some places, where
the effort of conforming to Guidelines is just too big for the real
gain. But I think the way this is done should be much more subtle,
based on an analysis of individual tradeoffs.
In your proposal the ring of the package is the important thing. But
the dependency tree is a very well connected graph, and various fringe
packages are used by "core" packages, so there's no way to develop
anything using just the core, and any real-life server is also going
to include packages from all rings. Security bugs in "fringe"
web-facing services are more important than bugs in "core" tools
which are purely local. For example, recent XSS vulnerability in
jquery [1] nicely cuts through all rings. Unbundling jquery here would
give a huge benefit, and does not become less important the farther a
package is from the core.
[1]
https://bugzilla.redhat.com/show_bug.cgi?id=1166041
== Premise ==
So, some time ago, we started talking about dividing up the Fedora
package set into a series of "rings". One of the driving purposes behind
this idea was to re-evaluate our policies for packaging in each of these
rings. A fair amount of time has passed and no formal discussions have
taken place (that I have seen, at any rate), so I'm going to provide
here a simple "starter" proposal that we can work from and see where it
leads.
To begin, I'll try to identify some of the major advantages and
disadvantages in our current policies. (In no particular order...)
=== Advantages ===
* Consistency: every shared library, language-specific module and
application should have packaging consistent with others in its
category. This makes it easier for comaintainers and provenpackagers to
pick it up and work on it.
* The no-bundled-libraries policy means that when a library module
requires an update, it means that only one package needs to be modified
in order to enhance all applications on the system that consumes it.
This is a significant time-saver when it comes to dealing with
(increasingly common) security vulnerabilities.
* Package review "front-loads" the task of educating the packagers. It
guarantees that packages enter the collection by meeting a very high
standard. This does a great deal to accomplish the "consistency"
mentioned above.
* Legal review and the FPCA ensures that Fedora remains true to its
"Freedom" Foundation (as well as protecting Fedora contributors from the
perils of the international legal system).
=== Disadvantages ===
* Very strict policies often leads to potential packagers giving up.
True. But
packagers often give up for many other reasons, including
unresolved legal problems with the software, many different packaging
problems other than bundling, lack of sponsors, etc.
* Package reviews for less-interestingackages (such as those for
less
popular SIGs) often remain un-reviewed for weeks, months or even years.
* The package policies are only ever reviewed during the initial
creation of the package. Once that initial (high) hurdle is cleared, the
packager essentially has free reign to do whatever they want with their
package. This sometimes means that as time passes, the spec files
"bit-rot" or otherwise start accumulating other inconsistencies. (A
common example is the package whose upstream starts bundling a library
without the packager noticing).
Your proposal would only worsen, not fix that.
* Many upstream projects do not concern themselves with being
"in" any
particular distribution (with the notable example being the
Debian/Ubuntu flavors which have amassed a sufficient apparent userbase
that they sometimes get special treatment). For a variety of reasons,
this often leads directly to bundling the vast majority of their
dependencies. This is done for many reasons, but the two most common are
supportability and portability; it's impossible for many upstreams to
actually QA their package with every possible distro library. Instead,
if they ship everything they depend on, they can guarantee *that*
specific combination. This moves the responsibility from the
distribution to the upstream package to maintain their bundled
libraries.
I think this responsibility is an illusion. Most upstreams only care
that the bundled versions allow their package to continue to work, and
are not at all capable of fixing vulnerabilities or even other bugs.
I think that it is also very unlikely that packages would be properly
reviewed when they move to an inner ring. Look at how well the merge
reviews are going.
* With many languages, there is no possibility of installing
multiple
versions of the same library on the same system, so if an application
requires a newer or older version of the library than is in Fedora to
run, it fails. For those languages that *do* allow this, it still adds a
significant maintenance burden on the packagers to keep multiple
versions of the same package up-to-date (which gets particularly hard
when upstream abandons a version that something else in Fedora depends
on...)
In practice, quite often it turns out that the version dependency
is not as strict as upstream makes it out to be. Recent example:
ElasticSearch [2]: despite initial concerns, it seems that ElasticSearch
will not require custom versions of any packages. Another
example: dateutil update [3]. Out of the many packages declaring
a dependency on dateutil 1.5, about 50% have been checked, and
they all work just fine with dateutil 2.4.
[2]
https://bugzilla.redhat.com/show_bug.cgi?id=902086
[3]
https://bugzilla.redhat.com/show_bug.cgi?id=1126521
== Proposal ==
With these things in mind, I'd like to propose that we amend the
packaging policy by splitting it into two forms:
=== Core Packages ===
Any package that is provided on a release-blocking medium (which at
present includes Fedora Atomic, Fedora Cloud, Fedora Server, Fedora
Workstation, the KDE Spin and several ARM images) must comply exactly
with the packaging guidelines as they are written today. These packages
must receive a full package review *when they are added to the install
media*. Any package present on the media if this proposal is adopted is
exempted from this requirement. Any package to newly appear on the
install media after this time *should* (I hate that word...) receive a
new package review, even if it is already present in Fedora.
=== Ring Packages ===
Any new package that is *not* going to be part of the install media set
is required to pass a lighter review and is permitted to carry bundled
libraries, with caveats to be listed below.
==== Ring Package Review Requirements ====
* A reviewer *MUST* establish suitability for the Fedora Project. This
means verifying that the package contains only permissible content
(legally-speaking).
* The package *MUST* conform to the FHS and other filesystem conventions
used by Fedora (such as %{python3_sitelib}), except when granted an
exception by the FPC.
* The package *MAY* contain bundled libraries or other projects, but if
it does so, it *MUST* contain a "Provides: bundled(pkg) = version" for
each such bundling. This is done so that we can use the meta-data to
identify which packages may be vulnerable in the event of a security
issue.
I'd like to make a counterproposal: analyze what Packaging Guidelines
cause the most problems, and periodically check if they are still worth
the effort. In specific cases, allow the package to be approved, but
open a bug against the package to keep the issue open and hopefully fix
it in the future.
Requirements which I'd drop right now:
1. strict adherence to FHS. There's only one /usr now, so the split
between /usr/share, /usr/lib, /usr/libexec is obsolete.
2. forbidding JNI libs in %{_libdir}. This is very popular, requires
modifying the source of the package to fix, and provides little
benefit. Instead it would be enough to require the JNI .so name to
not conflict. For example suffixing with 'j' would be enough.
I'm sure there are other things like that, which are pain in the ass
to fix, and could be dropped without loss to the distribution, in
contrast to a blanket exception to the guidelines for a large subset
of packages.
Zbyszek