Debian “testing” distribution

For basic, user-oriented information about the testing distribution, please see the Debian FAQ.

An important thing to note, both for regular users and the developers of testing, is that security updates for testing are not managed by the security team. For more information please see the Security Team's FAQ.

This page primarily covers the aspects of testing important to Debian developers.

How testing works

The testing distribution is an automatically generated distribution. It is generated from the unstable distribution by a set of scripts which attempt to move over packages which are reasonably likely to lack release-critical bugs. They do so in a way that ensures that dependencies of other packages in testing are always satisfiable.

A (particular version of a) package will move into testing when it satisfies all of the following criteria:

  1. It must have been in unstable for 10, 5 or 2 days, depending on the urgency of the upload;
  2. It must be compiled and up to date on all architectures it has previously been compiled for in unstable;
  3. It must not have release-critical bugs which do not also apply to the version currently in testing (see below for more information);
  4. All of its dependencies must either be satisfiable by packages already in testing, or be satisfiable by the group of packages which are going to be installed at the same time;
  5. The operation of installing the package into testing must not break any packages currently in testing. (See below for more information.)

A package which satisfies the first three of the above is said to be a Valid Candidate.

The update script shows when each package might move from unstable into testing. The output is twofold:

Frequently Asked/Answered Questions

What are release-critical bugs, and how do they get counted?

All bugs of some higher severities are by default considered release-critical; currently, these are critical, grave and serious bugs.

Such bugs are presumed to have an impact on the chances that the package will be released with the stable release of Debian: in general, if a package has open release-critical bugs filed on it, it won't get into testing, and consequently won't be released in stable.

The testing bug count are all release-critical bugs which are marked to apply to package/version combinations that are available in testingfor a release architecture.

How could installing a package into testing possibly break other packages?

The structure of the distribution archives is such that they can only contain one version of a package; a package is defined by its name. So, when the source package acmefoo is installed into testing, along with its binary packages acme-foo-bin, acme-bar-bin, libacme-foo1 and libacme-foo-dev, the old version is removed.

However, the old version may have provided a binary package with an old soname of a library, such as libacme-foo0. Removing the old acmefoo will remove libacme-foo0, which will break any packages which depend on it.

Evidently, this mainly affects packages which provide changing sets of binary packages in different versions (in turn, mainly libraries). However, it will also affect packages upon which versioned dependencies have been declared of the ==, <= or << varieties.

When the set of binary packages provided by a source package change in this way, all the packages that depended on the old binaries will have to be updated to depend on the new binaries instead. Because installing such a source package into testing breaks all the packages that depended on it in testing, some care now has to be taken: all the depending packages must be updated and ready to be installed themselves so that they won't be broken, and, once everything is ready, manual intervention by the release manager or an assistant is normally required.

If you are having problems with complicated groups of packages like this, contact debian-devel or debian-release for help.

I still don't understand! The testing scripts say that this package is a valid candidate, but it still hasn't gone into testing.

This tends to happen when in some way, directly or indirectly, installing the package will break some other package.

Remember to consider your package's dependencies. Suppose your package depends on libtool, or libltdlX. Your package won't go into testing until the right version of libtool is ready to go in with it.

In turn, that won't happen until installing libtool doesn't break things already in testing. In other words, until all other packages which depend on libltdlY (where Y is the earlier version) have been recompiled, and all their release critical bugs are gone, etc, none of these packages will enter testing.

This is where the textual output [gzipped] is useful: it gives hints (albeit very terse ones) as to which packages break when a valid candidate is added to testing (see the Developer's Reference for more details).

Why is it sometimes hard to get Architecture: all packages into testing?

If the Architecture: all package is to be installed, it must be possible to satisfy its dependencies on all architectures. If it depends on certain packages which only compile on a limited set of Debian's architectures, then it can't do that.

However, for the time being, testing will ignore Architecture: all packages' installability on non-i386 architectures. (It's a gross hack and I'm not really happy to have made it, but there you go. —aj)

My package is stalled because it's out of date on some architecture. What do I do?

Check the status of your package in the build log database. If the package doesn't build, it will be marked as failed; investigate the build logs and fix any of the problems that are caused by your package's sources.

If you happen to notice that some architectures have built the new version of your package, but it isn't showing up in testing scripts output, then you just have to be a bit more patient until the respective buildd maintainer uploads the files to the Debian archive.

If you notice that some architectures haven't built your new version of the package at all, despite the fact you uploaded a fix for an earlier failure, the reason is probably that it's marked as waiting for dependencies (Dep-Wait). You can also see the list of these so-called wanna-build states to make sure.

These problems usually get fixed eventually, but if you've been waiting for a longer period of time (say, two weeks or more), notify the respective port buildd maintainer if such an address is documented on the port web page, or the mailing list of the port.

If you have explicitly dropped the architecture from the Architecture list in the control file, and the package has been built for that architecture before, you will need to request that the old binary package for this architecture be removed from the archive before your package can transition to testing. You need to file a bug against ftp.debian.org requesting removal of the dropped architecture's packages from the unstable archive. Generally the relevant porting list should be informed as a matter of courtesy.

Are there any exceptions? I'm sure acmefoo has just made it into testing despite not satisfying all of the requirements.

The release manager can override the rules in two ways:

Can you provide a real, non-trivial example?

Here's one: when the source package apache is installed into testing, along with its binary packages apache, apache-common, apache-dev and apache-doc, the old version is removed.

However, all Apache module packages depend on apache-common (>= something), apache-common (<< something), so this change breaks all of those dependencies. Consequently, all Apache modules need to be recompiled against the new version of Apache in order for testing to be updated.

Let's elaborate on this a bit further: after all of the modules have been updated in unstable to work with a new Apache, the testing scripts try apache-common and find out that it breaks all the Apache modules because they have Depends: apache-common (<< the current version), and then try libapache-foo to find out that it doesn't install because it has Depends: apache-common (>= the new version).

However, later they'll apply a different logic (sometimes prompted by a manual intervention): they'll ignore the fact apache-common breaks stuff, and keep going with things that work; if it still doesn't work after we've done everything we can, too bad, but maybe it will work. Later they'll try all the random libapache-foo packages and see that they indeed work.

After everything's been tried, they check how many packages have been broken, work out if that's better or worse than what there was originally and either accept everything or forget about it. You'll see this in update_output.txt on recur: lines.

For example:

         recur: [foo bar] baz

basically says having already found that foo and bar make things better, I'm now trying baz to see what happens, even though that breaks things. The lines of update_output.txt that start with accepted indicate things that appear to make things better, and skipped lines make things worse.

The update_output.txt file is completely unreadable!

That is not a question. ;-)

Let's take an example:

 skipped: cln (0) (150+4)
     got: 167+0: a-40:a-33:h-49:i-45
     * i386: ginac-cint, libginac-dev

This means that if cln goes into testing, ginac-cint and libginac-dev become uninstallable in testing on i386. Note that the architectures are checked in alphabetical order and only the problems on the first architecture with problems are shown — that's why the alpha architecture is shown so often.

The got line includes the number of problems in testing on the different architectures (until the first architecture where a problem is found — see above). The i-45 means that if cln would go into testing, there would be 45 uninstallable packages on i386. Some of the entries above and below cln show there were 43 uninstallable packages in testing on i386 at the time.

The skipped: cln (0) (150+4) line means that there are still 150 packages to go through after this package until this check of all packages is completed, and that 4 packages have been found already that won't be planned to be upgraded because they would break dependencies. The (0) is irrelevant, you can safely ignore it.

Note that there are several checks of all packages in one testing script run.

Jules Bean initially assembled the frequently asked questions and answers.

Additional information

The following pages provide additional information about the current state of testing and the migration of packages from unstable to testing:

You might be interested in reading an older explanation email. Its only major flaw is that it doesn't take the package pool into account, because that was implemented by James Troup after it was written.

The testing code is available from ftp-master.

Anthony Towns takes credit for the implementation of testing.