The Annotated Lancaster Consensus

The official Lancaster Consensus document is on Github. This blog post is an annotated review of it.

The Lancaster Consensus

At the first Perl QA Hackathon in 2008 in Oslo, a number of QA and
toolchain authors, maintainers and experts came together to agree on some
common standards and practices. This became known as
"The Oslo Consensus".

Five years later, at the 2013 Perl QA Hackathon, a similar brain trust came
together to address new issues requiring consensus.

These decisions provide direction, but, as always, the speed of
implementation will depend on the interests and availability of volunteers
to do the work.

Toolchain and testing

Minimum-supported Perl

Going forward, the Perl toolchain will target Perl 5.8.1, released
September 2003. This will allow toolchain modules to reliably use PerlIO
and improved Unicode support.

Because of the many Unicode bug-fixes in early 5.8 releases,
toolchain maintainers reserve the right to later bump the
minimum to 5.8.4 (which ships with Solaris 10).

There was huge agreement about 5.8.1, and the 5.8.4 discussion hinged on the amount of work to avoid early 5.8 bugs compared to the number of people affected, particularly given that there is CPXXXAN to support older Perls.

Specifying pure-perl builds

Some distributions offer an "XS" version or a "Pure Perl" version that can
be selected during configuration. Currently, each of these has their own
way for users to indicate this, which makes it impossible for CPAN clients
or other build tools to help users select automatically.

For example, the module both checks the PERL_ONLY environment variable and the following three command line flags: --perl-only, --perl_only and --xs. But, the Sentinel module checks only for the command line flag --pp. Other modules do it yet differently, with different environment variables or command line options. It's chaos.

Going forward, the "spec" for Makefile.PL and Build.PL will include command
line options to request a "pure Perl only" build. These will be:

  • PUREPERL_ONLY=1 (for Makefile.PL)
  • --pureperl-only (for Build.PL)

These may be set in the PERL_MM_OPT or PERL_MB_OPT environment
variables just like any other command line option.

If present, distribution authors must ensure that the installed modules do
not require loading XS (whether directly or via Inline) or dynamically
generate any platform-specific code. The installed files must be able to
run correctly if copied to another machine with the same Perl version but a
different architecture (e.g. "fatpacking" an application). If this
condition can not be met, configuration must exit with an error.

Fatpacking is the most common use case for this, so module authors should think about that explicitly when deciding if they can be "pure perl only" or not.

Environment variables for testing contexts

The Oslo Consensus defined two testing contexts: AUTOMATED_TESTING and
RELEASE_TESTING. Of these, AUTOMATED_TESTING has been the most
confusing, as it sometimes was used to mean "don't interact with a user"
and sometimes "run lengthy tests".

I've also used it for tests which depended on some external website working correctly. I wouldn't want to stop someone from installing the module if it failed, but I did want to see what CPAN smokers experienced.

We also (briefly) discussed how some tools like Dist::Zilla are using

Distribution authors should now follow these semantics:

  • AUTOMATED_TESTING: if true, tests are being run by an automated testing
    facility and not as part of the installation of a module; CPAN smokers
    must set this to true; CPAN clients must not set this

  • NONINTERACTIVE_TESTING: if true, tests should not attempt to interact
    with a user; output may not be seen and prompts will not be answered

  • EXTENDED_TESTING: if true, the user or process running tests is willing
    to run optional tests that may take extra time or resources to complete.
    Such tests must not include any development or QA tests. Only tests of
    runtime functionality should be included.

  • RELEASE_TESTING: if true, tests are being run as part of a release QA
    process; CPAN clients must not set this variable

  • AUTHOR_TESTING: if true, tests are being run as part of an author's
    personal development process; such tests may or may not be run prior to
    release. CPAN clients must not set this variable. Distribution
    packagers (ppm, deb, rpm, etc.) should not set this variable.

AUTHOR_TESTING was not really discussed, but I included it in the writeup for completeness. It was discouraged in the Oslo Consensus, but some people seem to have tests they want run throughout development and others they want run only at release time, so it still gets used. For example, Dist::Zilla sets it for "dzil test" since that command is only run by authors, not by end users.

There are already two libraries on CPAN to make it easier to set these
variables correctly:

CPAN smokers and integration testers must indicate automated,
non-interactive testing and may request extended testing, depending on
their resources.

For example, a CPAN tester may decide not to run extended testing on old, slow hardware.

CPAN clients are free to request non-interactive or extended testing
depending on their needs or configuration.

CPAN smokers and clients that "must not set" a variable also must not clear
it if it is already set externally.

Amendments to the Build.PL spec

David Golden and Leon Timmermans have been working on a
Build.PL spec
to describe how any Perl build tool using Build.PL must behave. It is
necessarily based on Module::Build, but does not need to follow its
behaviors exactly.

The group agreed that the use and semantics of .modulebuildrc should
be excluded from the specification.

Installed distributions database

One of the QA hackathon projects was the creation of a replacement
for packlists. An installed-distribution database would facilitate
easy inventory of installed distributions, uninstall tools and tracking of
the dependency graph of installed modules.

The consensus discussions were explicitly not designing the system; the brief was to answer questions about the various ways/places modules can be installed so people doing the actual design work didn't paint themselves into a corner.

The group agreed that because modules can be installed into many different
locations, any such database would need to be "per @INC" and that it would
need to stack in the same way that @INC itself does. That means that
adding paths to @INC could change what the database sees as installed.

Such a database system must not require any non-core dependencies, but
could offer enhanced capabilities if recommended CPAN modules are

Other implementation details are left to anyone designing such a system.

Post-installation testing

Several people at the hackathon have been interested in a system for
running module tests after installation, for example to ensure that
upgraded dependencies don't break a module or to test overall integrity.

The group agreed that any such testing must make all distribution files
available during testing -- tests must be run from within a distribution
tarball directory. Any such tests must be run using new make or
Build targets: make test-installed or Build test-installed. These
should be equivalent to make test or Build test but without adding
blib to @INC. The prove application must not be used.

These targets don't exist and would have to be created in each tool. But conceptually they should work just like "make test" would, except they should run against the installed modules, not the ones built into "blib".

The group also agreed that any such tests need to respect how modules can
be shadowed in @INC. Setting PERL5LIB could change which is the
"installed" distribution and thus which tests should run. Coordination
with an installed distribution database was encouraged.

Other implementation details, including whether the distribution directory
is saved from the initial installation or retrieved fresh from CPAN/BackPAN,
are left to anyone designing such a system.

META file specification

The 'provides' field

The 'provides' field of the
CPAN::Meta::Spec requires a 'file'
sub-key, but the meaning was unclear for dynamically-generated packages.
We agreed that the 'file' key must refer to the actual file within the
distribution directory that originates the package, whether that is a .pm
file or a .PL or other dynamic generator.

The group also agreed that having a required 'file' sub-key didn't make sense, but I realized afterwards that changing the spec would break any existing validators and that chaos that could cause wouldn't be worth the benefit. But it's absolutely worth making it optional for whenever we get to v3 of the spec.

Improving on 'conflicts'

We briefly discussed some of the known problems with the 'conflicts' key
within prerequisite data.

What most developers seem to want is a way to indicate that installing a
particular module is know to break other modules of particular versions.
E.g. upgrading Foo to 2.0 breaks any Bar before 3.14.

We encouraged anyone interested in improvements to prototype it using an
x_breaks or similar custom key and getting patches to support it into
CPAN clients. Once battle tested, it could be a candidate for a future v3
of the spec.

This discussion had huge risk of turning into a design discussion, so we declared that people should prototype with a custom key rather than get into a spec discussion prematurely.


Long-term goal for distribution-level data on PAUSE

Several of the PAUSE issues discussed highlighted the need for PAUSE to
maintain not just package (namespace) level index and permission data, but
also "distribution" level data. This would allow, for example,
transferring permissions for a distribution as a unit instead of needing
to transfer permissions on all packages.

We agreed that this is the right long-term goal, but that other proposals
would be implemented in the near-term to solve current issues.

This was a classic "good", "fast" and "cheap" tradeoff. With volunteer labor, we are "cheap". The long-term idea was "good", but we agreed that we wanted something "fast" more than we wanted something "good" so the rest of the proposals represent what could be done quickly.

Case insensitive package permissions

While not discussed directly, it should be noted that PAUSE package
permissions will shortly become case-insensitive, but case-preserving
to ensure that indexed modules would be unique even if installed on a
case-insensitive file system.

For example, there was a File::Stat on CPAN. Installing it into sitelib on a case-insensitive system (like Mac OS X), meant that use File::stat would actually load File::Stat. The core module would be completely hidden. Ouch!

Rules for distribution naming

Many CPAN ecosystem websites and tools treat a "distribution name" as
a unique identifier, even though nothing has enforced uniqueness to date.
Allowing non-uniqueness is confusing at best and a security risk at worst.

Gory details are in this email to "Distribution names are not unique..."

Going forward, distributions uploaded to PAUSE must have a name that
"matches" the name of an indexed package within the distribution and the
uploader must have permissions for that package or else the entire
distribution will not be indexed.

For example, if DAGOLDEN uploads Foo-Bar-1.23.tar.gz, the distribution name
is "Foo-Bar" and there must be an indexable "Foo::Bar" package within the

There are about 1000 distributions on CPAN that do not follow this rule and
they will be grandfathered, though they are encouraged to conform to the
standard either by renaming the distribution, adding a new .pm file or by
introducing a properly named package internally.

For example, LWP ships as libwww-perl-6.05.tar.gz. If it included package
into one of its .pm files, that package would be indexed and
would conform with the standard.

Technically, the correct package could also be declared only in the
META.json file using a 'provides' field. In such a case the 'file' sub-key
must be 'META.json' to indicate that 'META.json' is the file responsible
for declaring the package.

Flagging abandoned modules and modules requesting help

Currently, when a CPAN author passes away, his or her module permissions
are transferred to a fake author called 'ADOPTME'. Volunteers can step
up to request a takeover if they wish to maintain them.

We agreed that in the short-term, a similar mechanism should be used to
signal abandonment or that an author is looking for someone to share
responsibility. Unlike the case where an author is deceased, these will
use co-maint privileges as a signaling mechanism so that the original
author may remove them as needed.

(In the long-term, the group hopes that a distribution-level data model for
PAUSE will be able to address these needs more directly.)

CPAN search engines and other community sites may use these permissions
markers and associated meanings to communicate the status of distributions.

  • ADOPTME as primary: this generally indicates a deceased author.
    Volunteers can request a takeover via

  • ADOPTME as comaint: this indicates a verified, non-responsive author.
    The community may propose that a package be so marked following the same
    rules as for a take-over (i.e. multiple attempts to contact the author
    and a request via Volunteers can request a takeover
    of an ADOPTME module via without an additional waiting

  • HANDOFF as comaint: this indicates that an author wishes to
    permanently give up the primary maintainer role to someone else

  • NEEDHELP as comaint: this indicates that an author seeks people to
    help maintain the module, but plans to continue as primary maintainer

It's very important that CPAN search engines treat ADOPTME differently from HANDOFF or NEEDHELP. Flagging one's module as "NEEDHELP" shouldn't result in a big red "Abandoned module!" warning.

Matt S. Trout has voluntered to administer requests for modules to be flagged as co-maint ADOPTME. Proposals must follow the normal rules for takeover. You must make several public, documented attempts to contact the author publicly before appealing to for ADOPTME to get comaint.

With the exception of a 'takeover' from ADOPTME (which must go through, CPAN authors must manage these comaint privileges using
the regular PAUSE interface.

A "takeover" from ADOPTME can be immediate because PAUSE admins already know that the author is non-responsive for whatever reason.

An author may also voluntarily transfer primary or co-maint to ADOPTME to
indicate that PAUSE admins may transfer permissions immediately to anyone
who requests it.

Automating PAUSE ID registration

Historically, PAUSE ID's have been manually approved, often with a
substantial delay. We agreed that assuming appropriate protections against
bots/spam are in place, PAUSE should move to an automated approval system.
This would bring it in line with other programming language repositories
and open source community sites.

Additionally, we agreed that unused, inactive PAUSE IDs should be deleted
and made available for reuse after a period of time. Specifically, any
PAUSE ID that ever uploaded anything must not be deleted (because the files
exists on BackPAN under that PAUSE ID). A login to PAUSE (or via a proxy
like is sufficient to indicate activity. Inactive IDs will
not be deleted without a warning message about logging in to PAUSE.

Automating CPAN directory cleanup

Approximately half the files on CPAN are older than 5 years. Many authors
never clean up old distributions. In order to keep the size of CPAN down,
we agreed that under certain conditions, old distribution will be
automatically scheduled for deletion (and will thereafter only exist on

For a distribution to be selected for deletion, there must be at least 3
stable releases. Anything older than the oldest of those 3 will be
scheduled for deletion if it is older than 5 years and is not indexed in
the 02packages file.

This is a bit confusing, but is intended to be really conservative. For example, if I have Foo-Bar-1.24, Foo-Bar-1.23_03, Foo-Bar-1.23_02, Foo-Bar-1.23_01, Foo-Bar-1.22, Foo-Bar-1.21_01, Foo-Bar-1.20 and Foo-Bar-1.18, Foo-Bar-1.20 is the third oldest stable release, so only Foo-Bar-1.18 would be considered for deletion if more than 5 years old and not indexed in 02packages. The 1.23_XX and 1.21_XX dev releases will be kept.

All perl tarballs will be excluded from deletion, of course.

Scheduled deletion will notify the author as usual and they will have the
usual period of time to cancel the scheduled deletion.

Cleanup will be implemented on some sort of rolling basis by author ID to
avoid bothering authors with frequent deletion notices.

Module registration

The group agreed that the PAUSE module registration has largely outlived its
usefulness. Because only a fraction of CPAN modules are registered,
registration does not provide a comprehensive source of metadata (e.g.
"DSLIP") and much of the information registration covers is more widely
available via META files.

One benefit if module registration is that the data can be changed without requiring a new release the way META files do. On reflection, about the only field that matters for is the "support level".

The group acknowledged the remaining benefit has been that new CPAN authors
often attempt to register their first module and benefit from feedback, but
felt that other venues, such as PrePAN, would offer a
better new author experience. In particular, PrePAN offers community
participation beyond one or two PAUSE admins and a wealth of examples to
learn from (without having to search through a mailing list archive).

brian d foy has been the module registration hero, tirelessly responding to requests for years. PrePAN will help share the burden and new authors will benefit from different points of view.

Therefore, we agreed that existing PAUSE documentation will be changed to
direct new (and experienced) authors to PrePAN for guidance.

Soon, PAUSE will stop publishing the module registration database to CPAN
mirrors. (The index file will exist but be empty to avoid breaking CPAN
clients that expect it.) After an assessment period, module registration
will likely be closed and this feature will be retired from PAUSE.

Participants in the Lancaster Consensus discussions

Discussions lasted over 3 days, participants came and went, but each day
had about 20 people. Thank you to the following participants:

Andreas König, Barbie, Breno Oliveira, Chris Williams, Christian Walde,
David Golden, Daniel Perrett, Gordon Banner, H. Merijn Brand, James
Mastros, Jens Rehsack, Jess Robinson, Joakim Tormoen, Kenichi Ishigaki,
Leon Timmermans, Liz Mattijsen, Matthew Horsfall, Michael Schwern, Olivier
Mengué, Paul Johnson, Peter Rabbitson, Philippe Bruhat, Piers Cawley,
Ricardo Signes, Salve J. Nilsen and Wendy van Dijk

(Apologies to anyone present who was left off the list. Email dagolden at
cpan dot org or send a pull request to be added.)

This entry was posted in cpan, perl programming, toolchain and tagged , , , , , . Bookmark the permalink. Both comments and trackbacks are currently closed.


  1. Joel Berger
    Posted April 19, 2013 at 2:35 pm | Permalink

    I have two quick points.

    Re: "Environment variables for testing contexts" I would like some mechanism for testers to indicate that they are testing without installing, this is important for modules that are part of the toolchain like Alien::Base and Module::Build::CleanInstall. An example are testers that append just-built blib directories to @INC which is very problematic for Alien::Base particularly on Mac and other platforms that use Mach (yes h) binaries.

    Re: "Automating CPAN directory cleanup" I can see the need to mark a dist/version as do-not-delete. This would be most needed for the case when the next version is breaking and that version remains for continued compatibility.

    • Posted April 19, 2013 at 8:18 pm | Permalink

      I think for your testing issue, that's so specific to CPAN Testers that you can just take that up on the CPAN Testers Discussion mailing list. There are only really two smoke platforms, so a patch along those lines for each is likely to be well-received. I'm not sure what you would do with that information, but I see no harm in providing it.

      For the cleanup, nothing indexed would be deleted. And we've said we'd keep 3 stables. So even if there is a new release and it's broken and the last release was 6 years ago, it wouldn't be deleted.

  2. Posted April 19, 2013 at 5:31 pm | Permalink

    How is x_breaks different than the existing conflicts key? If upgrading Foo to 2.0 breaks Bar before 3.14, then `conflicts => { Bar => '3.14' }` is exactly right... so I'm not sure of the distinction that is being made between the two.

    • Posted April 19, 2013 at 8:20 pm | Permalink

      Conflicts doesn't really have any defined meaning. I think the possible distinction is that "conflicts" means "don't install me if this other thing is there" and "x_breaks" or whatever means "if you upgrade me, also upgrade these other things if you can". But until someone actually implements something along those lines, the semantics aren't clear, which is why we said "go prototype it with x_whatever". If it catches on and is useful, the we can consider making it "official" in a future v3 spec.

      • Posted April 20, 2013 at 1:00 pm | Permalink

        I'm in agreement if the meaning of these two fields were reversed -- as x_conflicts has already been in use for some time (since at least November 2010, when Dist::CheckConflicts was first released) to mean "if you upgrade me, these other things will be broken and need to be upgraded". For example, see .

        • Posted April 20, 2013 at 11:51 pm | Permalink

          Well, you can call it whatever you want if it starts with "x_". Personally, I think "x_conflicts" is confusing because it doesn't mean quite what "conflicts" in a prerequisite section does.