Friday, 5 April 2019

Fixing proposed-migration and autopkg tests (Re: Maintaining language-specific module package stacks)

On 03.04.19 13:41, Robie Basak wrote:
> I'd like to talk about addressing the difficulty in maintenance of long
> tail language-specific stacks in Ubuntu. For example, right now
> `src:rails` is stuck in disco-proposed[1]. It seems to me that we spend
> a disproportionate amount of effort trying to get this class of package
> migrated to the release pocket compared to the number of Ubuntu users
> who actually care and use them.

We certainly can remove some packages to reduce the workload on proposed
migration. But then we can remove the complete universe component except for
build dependencies for saving even more effort (how will that affect the
popularity of Ubuntu?). Instead please can we fix the effort we have to spend
on proposed migration?

The introduction of the proposed pocket and the britney migration to the release
pocket did solve a lot of issues, making the release pocket of development
releases more stable. Migrations to the release pocket usually were only
blocked by builds failures (and test failures when tests were run as part of the
build). I assume that migration blockers were well understood with the help of
transition trackers. When doing a transition of a package in main, a developer
still had to address all the no-change rebuilds. With the introduction of
autopkg tests this effort increases a lot, maybe even unforeseen when
introducing the autopkg tests.

Our autopkg test infrastructure is too unreliable:

(1) A percentage of fails are shown with the /unknown version.
These usually go away when given back, and this can be done
automatically. I think somebody is doing this on a regular basis.
How often, is this documented?

(2) When the autopkg test queues are full, you still see failures
besides the ones in (1), and which go away when you re-run the
tests when the queues are empty. These failures are typically
timeouts, or in packages which may have a big I/O load. These
give-backs are done manually, and delay a migration until the
queues are empty. We currently don't have any means to track
these issues.

(3) Build and test times in autopkg tests vary up to a factor of five,
sometimes leading to timeouts. Some packages are known, but
there is no means to debug these (and why should an uploader do

Ideally an autopkg test should have the resources to succeed with at the first
try. Any retry is extra effort.

(4) It looks like the failures in (2) and (3) are caused by running
under I/O and CPU constraints, but we don't have any means to
validate that, nontransparent to the developer trying to
migrate something. Most autopkg tests are tests running the
upstream tests of a package, and these are usually functional
tests, and maybe not designed to run and succeed under load
conditions. It is a conflicting goal to run as many tests as
possible and to provide a test with sufficient resources. This
problem will only aggravate because the number of autopkg tests
to run increases non-linearly with each autopkg test added,
and there are plans to trigger autopkg tests for indirect
dependencies as well...


> I suggest that:
> 1. If a language-specific package, or stack of packages, is stuck in
> proposed, and nobody is volunteering to get them migrated, then we are
> more willing to delete them from the release pocket and release without
> that stack.
> 2. We recommend, as a project, that users who wish to use these
> language stacks directly do so via the language-specific packaging
> tooling.
> Background
> ----------
> It'd be great to get input from language-specific communities who may
> not be intimately familiar with distribution development process, so
> here's a quick summary of what I'm talking about. Those already familiar
> with distribution development workflow can skip to the next section.
> You're presumably aware that language-specific communities generally
> have their own package repositories and package managers, such as
> PyPI/pip,, npm Registry/npm and so on. My understanding
> is that these communities generally advise that users consume from these
> repositories directly.
> Debian often packages a subset of these repositories - usually for the
> purpose of fulfilling the dependency requirement of some higher level
> component, such as Rails in my example. Ubuntu then autosyncs these.
> Some users prefer these higher level components to come through their
> distribution, rather than via a language-specific package manager -
> presumably for the release management consolidation that this provides.
> However, it's my understanding that in practice the majority of users do
> not consume the distribution packaging for these components, instead
> using the language-specific package managers directly as generally
> recommended by those upstreams. It also seems quite common for language
> communities to specifically recommend against consuming language module
> packages through the distribution.
> In general distributions include only one version of each component in a
> given distribution release. Bringing the entire dependency web in line
> to make this possible can take considerable effort on the part of
> distribution developers, particularly when upstreams tend to use a large
> number of small dependencies and also require specific dependency
> versions due to frequent API breaks.
> In practice, the dependency web is more complicated than this, and
> entire swathes of packages get held up at once until all the
> dependencies can be resolved together; this takes work in figuring out
> what the holdups are, making decisions about which versions to ship, and
> possibly patching code to change what dependency versions are actually
> required, in order to make everything work together. Think of it like
> trying to come up with a single `requirements.txt` (as generated by `pip
> freeze`), `Gemfile.lock` or `package-lock.json` file that works for all
> packages across the entire distribution release.
> Until this is resolved, distribution package updates are held in a
> staging area, and won't be part of the next distribution release. In
> Ubuntu we say that that our packages are "stuck in proposed". Once
> resolved, the packages "migrate" to the "release pocket".
> The details of this process in Ubuntu are documented here:
> For example, we are blocked from shipping a newer version of the PHP
> interpreter until the `wordpress` package (which is written in PHP)
> works with the proposed newer PHP version. Now consider that there are
> hundreds of these reverse dependency packages like `wordpress`,
> including many PHP language modules. The addition of each one makes it a
> little harder for Ubuntu to move on updating PHP itself. My suggestion
> is that we more readily remove these reverse dependencies from the
> distribution release to free up the update of PHP itself in this
> example, rather that spend what seems like a disproportionate amount of
> effort fixing things that we suspect very few users actually care about.
> Discussion
> ----------
> Part of the purpose of my suggestion is to make it far easier to
> transition to new language interpreters without worrying too much about
> the very long tail of barely used reverse dependency language modules
> that usually hold up these transitions. My understanding is that far
> more users rely on the distribution supplying the language interpreter
> and language-specific package manager than the modules themselves. In
> case of a transition being held up like this, I'm proposing to simply
> permit the deletion of the long tail from the release pocket, get the
> newer interpreter stack migrated into the release pocket, and consider
> it done. The long tail will then migrate if maintained actively by
> others, and if it isn't, we'll ship without it.
> Right now for example, I'm suggesting that we simply delete `src:rails`
> from the release pocket, including its reverse dependencies, unless the
> reverse dependency list contains something outside the Rails stack that
> is unacceptable to us to delete. Rails users who use `gem install
> rails`, as recommended by Rails upstream[2], will not be affected.
> My suggestion deliberately leaves the door open for able volunteers to
> be able to maintain these packages in Ubuntu if they wish.
> One downside to my suggestion is that availability of particular
> language module packages may become unreliable between Ubuntu releases
> from a user's perspective (if a particular package skips an Ubuntu
> release before being restored, for example). I suggest that this can be
> tackled later if it becomes a problem, for example by blacklisting such
> packages for longer unless a team is prepared to commit to preventing
> this from happening.
> Another potential downside is in packages that are generally useful to
> users outside their own language ecosystems, yet depend on these
> language-specific dependency stacks. For example, take Vagrant. Vagrant
> is written in Ruby, so needs packages originating from I'm
> sure there are plenty of users who aren't deploying a Ruby-based stack
> but who do use Vagrant. I think that a significant number of these users
> probably prefer to consume Vagrant from the distribution package (`apt
> install vagrant`) rather than from (`gem install
> vagrant`)[3]. Because distributions include all of their dependencies in
> their own repositories, this means that to ship Vagrant as a package in
> Ubuntu, we also need to ship a package for everything in
> that Vagrant requires. I suggest that we don't apply my deletion policy
> to such packages and their dependencies - that we continue trying to
> maintain them on a best effort basis as we do today.
> If we do change anything in this regard, I think an important part of
> this is for us to decide as a project what affected users can expect,
> what we recommend that they do, and that we communicate this clearly. I
> suggest that, in the absence of Ubuntu development teams volunteering to
> maintain this class of packages in Ubuntu, we generally follow
> upstream's recommendations in using their language-specific package
> management stack rather than apt/dpkg. This isn't all that different
> from the situation today, where users can't rely on us having packaged
> the specific modules that they need at particular versions anyway.
> We might want to be more specific in our recommendations to users. For
> example, it generally is better for system stability, particularly when
> installing software from third party sources, for third party software
> to be confined well. Where such a system is available, we could
> specifically recommend its use, and recommend against installing to the
> "system". For example, with the use of virtualenv for Python stacks. I
> note that Ruby is available from upstream as a snap, so if suitable for
> general deployments that might be the gold standard for recommended
> confinement; if not, then at least rvm.
> Please discuss. I intend to draw this thread to the attention of
> language-specific communities too, to try and get their input. In
> particular it'd be great to agree on the same recommendations for Ubuntu
> users. Note that this list is moderated; I propose to permit emails from
> language-specific communities in response to this thread for a few weeks
> to avoid fragmenting discussion between here and the unmoderated list.
> I'll make sure replies get through moderation promptly. Don't worry
> about not being subscribed: just replying to the list will be fine.
> Thanks,
> Robie
> [1]
> [2]
> [3] I note that Vagrant upstream specifically advises against `apt
> install vagrant` from distributions[4]. Contrary to my understanding of
> typical _direct_ use of language-specific stacks with which I'm
> justifying my suggestion, however, my understanding is that many of
> those who use Vagrant but aren't Ruby-based developers prefer the
> distribution package regardless. It makes sense to me that a
> non-Ruby-developer wouldn't want to learn, use and maintain an entirely
> different package manager or add an additional third party software
> source just for one top level package that the distribution ships
> anyway, unless they specifically really need a version newer than one
> provided by the distribution release.
> [4]

ubuntu-devel mailing list
Modify settings or unsubscribe at: