Working with buildbots#

To assert that there are no regressions in the development and maintenance branches, Python has a set of dedicated machines (called buildbots or build workers) used for continuous integration. They span a number of hardware/operating system combinations. Furthermore, each machine hosts several builders, one per active branch: when a new change is pushed to this branch on the public GitHub repository, all corresponding builders will schedule a new build to be run as soon as possible.

The build steps run by the buildbots are the following:

  • Check out the source tree for the changeset which triggered the build

  • Compile Python

  • Run the test suite using strenuous settings

  • Clean up the build tree

It is your responsibility, as a core developer, to check the automatic build results after you push a change to the repository. It is therefore important that you get acquainted with the way these results are presented, and how various kinds of failures can be explained and diagnosed.

In case of trouble#

Please read this page in full. If your questions aren’t answered here and you need assistance with the buildbots, a good way to get help is to either:

  • contact the mailing list where all buildbot worker owners are subscribed; or

  • contact the release manager of the branch you have issues with.

Buildbot failures on pull requests#

The bedevere-bot on GitHub will put a message on your merged Pull Request if building your commit on a stable buildbot worker fails. Take care to evaluate the failure, even if it looks unrelated at first glance.

Not all failures will generate a notification since not all builds are executed after each commit. In particular, reference leaks builds take several hours to complete so they are done periodically. This is why it’s important for you to be able to check the results yourself, too.

Checking results of automatic builds#

There are three ways of visualizing recent build results:

  • The Web interface for each branch at, where the so-called “waterfall” view presents a vertical rundown of recent builds for each builder. When interested in one build, you’ll have to click on it to know which changesets it corresponds to. Note that the buildbot web pages are often slow to load, be patient.

  • The command-line client, which you can get from Installing it is trivial: just add the directory containing to your system path so that you can run it from any filesystem location. For example, if you want to display the latest build results on the development (“main”) branch, type: -q 3.x
  • The buildbot “console” interface at This works best on a wide, high resolution monitor. Clicking on the colored circles will allow you to open a new page containing whatever information about that particular build is of interest to you. You can also access builder information by clicking on the builder status bubbles in the top line.

If you like IRC, having an IRC client open to the #python-dev-notifs channel on is useful. Any time a builder changes state (last build passed and this one didn’t, or vice versa), a message is posted to the channel. Keeping an eye on the channel after pushing a changeset is a simple way to get notified that there is something you should look in to.

Some buildbots are much faster than others. Over time, you will learn which ones produce the quickest results after a build, and which ones take the longest time.

Also, when several changesets are pushed in a quick succession in the same branch, it often happens that a single build is scheduled for all these changesets.


A subset of the buildbots are marked “stable”. They are taken into account when making a new release. The rule is that all stable builders must be free of persistent failures when the release is cut. It is absolutely vital that core developers fix any issue they introduce on the stable buildbots, as soon as possible.

This does not mean that other builders’ test results can be taken lightly, either. Some of them are known for having platform-specific issues that prevent some tests from succeeding (or even terminating at all), but introducing additional failures should generally not be an option.

Flags-dependent failures#

Sometimes, while you have run the whole test suite before committing, you may witness unexpected failures on the buildbots. One source of such discrepancies is if different flags have been passed to the test runner or to Python itself. To reproduce, make sure you use the same flags as the buildbots: they can be found out simply by clicking the stdio link for the failing build’s tests. For example:

./python.exe -Wd -E -bb  ./Lib/test/ -uall -rwW


Running Lib/test/ is exactly equivalent to running -m test.

Ordering-dependent failures#

Sometimes the failure is even subtler, as it relies on the order in which the tests are run. The buildbots randomize test order (by using the -r option to the test runner) to maximize the probability that potential interferences between library modules are exercised; the downside is that it can make for seemingly sporadic failures.

The --randseed option makes it easy to reproduce the exact randomization used in a given build. Again, open the stdio link for the failing test run, and check the beginning of the test output proper.

Let’s assume, for the sake of example, that the output starts with:

./python -Wd -E -bb Lib/test/ -uall -rwW
== CPython 3.3a0 (default:22ae2b002865, Mar 30 2011, 13:58:40) [GCC 4.4.5]
==   Linux-2.6.36-gentoo-r5-x86_64-AMD_Athlon-tm-_64_X2_Dual_Core_Processor_4400+-with-gentoo-1.12.14 little-endian
==   /home/buildbot/buildarea/3.x.ochtman-gentoo-amd64/build/build/test_python_29628
Testing with flags: sys.flags(debug=0, inspect=0, interactive=0, optimize=0, dont_write_bytecode=0, no_user_site=0, no_site=0, ignore_environment=1, verbose=0, bytes_warning=2, quiet=0)
Using random seed 2613169
[  1/353] test_augassign
[  2/353] test_functools

You can reproduce the exact same order using:

./python -Wd -E -bb -m test -uall -rwW --randseed 2613169

It will run the following sequence (trimmed for brevity):

[  1/353] test_augassign
[  2/353] test_functools
[  3/353] test_bool
[  4/353] test_contains
[  5/353] test_compileall
[  6/353] test_unicode

If this is enough to reproduce the failure on your setup, you can then bisect the test sequence to look for the specific interference causing the failure. Copy and paste the test sequence in a text file, then use the --fromfile (or -f) option of the test runner to run the exact sequence recorded in that text file:

./python -Wd -E -bb -m test -uall -rwW --fromfile mytestsequence.txt

In the example sequence above, if test_unicode had failed, you would first test the following sequence:

[  1/353] test_augassign
[  2/353] test_functools
[  3/353] test_bool
[  6/353] test_unicode

And, if it succeeds, the following one instead (which, hopefully, shall fail):

[  4/353] test_contains
[  5/353] test_compileall
[  6/353] test_unicode

Then, recursively, narrow down the search until you get a single pair of tests which triggers the failure. It is very rare that such an interference involves more than two tests. If this is the case, we can only wish you good luck!


You cannot use the -j option (for parallel testing) when diagnosing ordering-dependent failures. Using -j isolates each test in a pristine subprocess and, therefore, prevents you from reproducing any interference between tests.

Transient failures#

While we try to make the test suite as reliable as possible, some tests do not reach a perfect level of reproducibility. Some of them will sometimes display spurious failures, depending on various conditions. Here are common offenders:

  • Network-related tests, such as test_poplib, test_urllibnet, etc. Their failures can stem from adverse network conditions, or imperfect thread synchronization in the test code, which often has to run a server in a separate thread.

  • Tests dealing with delicate issues such as inter-thread or inter-process synchronization, or Unix signals: test_multiprocessing, test_threading, test_subprocess, test_threadsignals.

When you think a failure might be transient, it is recommended you confirm by waiting for the next build. Still, even if the failure does turn out sporadic and unpredictable, the issue should be reported on the bug tracker; even better if it can be diagnosed and suppressed by fixing the test’s implementation, or by making its parameters - such as a timeout - more robust.

Custom builders#

When working on a platform-specific issue, you may want to test your changes on the buildbot fleet rather than just on GitHub Actions and Azure Pipelines. To do so, you can make use of the custom builders. These builders track the buildbot-custom short-lived branch of the python/cpython repository, which is only accessible to core developers.

To start a build on the custom builders, push the commit you want to test to the buildbot-custom branch:

$ git push upstream <local_branch_name>:buildbot-custom

You may run into conflicts if another developer is currently using the custom builders or forgot to delete the branch when they finished. In that case, make sure the other developer is finished and either delete the branch or force-push (add the -f option) over it.

When you have gotten the results of your tests, delete the branch:

$ git push upstream :buildbot-custom     # or use the GitHub UI

If you are interested in the results of a specific test file only, we recommend you change (temporarily, of course) the contents of the buildbottest clause in; or, for Windows builders, the Tools/buildbot/test.bat script.

Ver também

New buildbot workers