Puppet pyup merges¶

CIDuty is currently assisting with pyup updates to the puppet repo. They regularly need help with these. This is a document to help releaseduty help CIDuty with these merges.

It may be a good idea to read through this document before making changes.

Get the PR into good shape¶

The first part is to get the PR in good shape. For these PRs especially, the taskcluster task results are important. These tests verify that the modules are all compatible with each other, with the caveat that the nodownload packages may be ignored.

Upstream pins that don’t match latest¶

For upstream pins that don’t match latest, we need to find the version that will work. For example, the test revealed the error

/build-puppet/modules/transparency_scriptworker/files/requirements.txt: taskcluster 5.0.0 has requirement taskcluster-urls<11,>=10.1.0, but you'll have taskcluster-urls 11.0.0 which is incompatible.


This is due to this line. To fix, we had to downgrade taskcluster-urls to 10.1.0.

We can either use the previous-good version, or we can search pypi for the package and look at the release history.

requirements files with stale ignores or pins¶

We may have pinned the upstream-of-upstream module in the puppet requirements files themselves. For example, if module1==1.5.0 requires module2<3.0, we may have pinned module2==2.0.1 with a # pyup: ignore or a # pyup: <3.0. However, when module1 bumps to a new version, that new version may require module2>=3.0. If we leave this, we may never be able to update either. If we catch this, we can try updating both module and module2.

When to land¶

Because of the nature of our puppet deployments, we may see hiccups in services during puppet deployment. For instance, if we update the signing_scriptworker requirements, when those instances puppetize, we may see failures in signing tasks.

We probably want to avoid landing these puppet updates during or right before chemspills or other critical releases. Other than that, we likely want to merge when we have time to monitor the puppet mail and help deal with fallout.

Landing¶

When we land, CIDuty will monitor treeherder and puppet mail. It may be a good idea for us to keep an eye on #ci and puppet mail as well. If we hit errors, papertrail can be helpful.

When we land a new change to the requirements files, puppet will download any missing modules. There is often a gap between 1) the first machines trying to puppetize with the new requirements files, and 2) the module being available on the puppet servers, so there is likely going to be a wave of puppet failures immediately following a pyup merge.

We need to make sure that these clear up within a puppet pass or two, or there may be something wrong.

For instance, sometimes a wheel is good for any python or platform. For example, urllib3-1.24-py2.py3-none-any.whl is good for python2 or python3, for any platform on any architecture. If the wheel contains compiled code, it may be more restrictive. For example, yarl-1.2.6-cp36-cp36m-manylinux1_x86_64.whl is compiled for cpython 3.6, linux, on 64bit intel type CPUs. If we’re on a different python (3.7, 3.5), a different platform (mac), or different architecture (arm or 32 bit), pip won’t use that wheel.

If pip is complaining about not finding a matching file, one way to debug is

• ssh releng-puppet2.srv.releng.mdc1.mozilla.com

• for python2, cd /data/python/packages; for python3, cd /data/python/packages-3.x

• look for the package name that’s missing. Note that pip will treat a dash - the same as an underscore _ in the package name, so python-jose is the same module as python_jose, for example. If we’re missing, say, the mac wheel (we only have a manylinux1 wheel for that module + version), we should download it:

• search for the module name on pypi

• click on the correct module, e.g. yarl

• if the latest version doesn’t match the version we want, find it in the release history, e.g. yarl release history; click on the correct version

• look at the download files for the wheel with the correct python version, platform, and arch. If there is no such wheel, go for the source tarball or zip.

I tend to download this into my homedir and move it into the correct repo, e.g.

pushd ~
wget https://...
cd /data/python/packages-3.x  # for 3.x packages; otherwise /data/python/packages
sudo mv -n ~/FILENAME .  # sudo allows me to write in the repo dir;
# mv -n doesn't overwrite an existing file


Puppet will rsync this file around. Once the correct file exists on any puppet master, pip should be able to find it.

There may be download problems if we point at a version or module that doesn’t exist on pypi. The fix here probably involves creating a new PR to point to the right module version, and merging it.

“stuck” installs¶

There may be a compile problem, a configuration problem, a dependency issue, or other.

First, is this affecting an entire class of machines, or just one machine? This is hard to tell when dealing with singleton pools, but if all signing scriptworkers are fine except for one, then it’s likely a machine issue. If all signing scriptworkers are throwing an error during puppetizing, then it’s probably a puppet issue.

Puppet’s pip logging can be a bit terse, but it shows you the command-line used. If you log into an affected machine and try to run that pip command manually, you may get a better sense of what’s wrong, possibly adding the -v option to be more verbose, e.g.

# You want to be su'ed as the user that owns /builds/scriptworker, probably cltbld or cltsign


If there’s a compile issue, we may need to grab a wheel or add system packages. We may want to delay this update at this time.

If there’s a dependency problem, we may need a new PR that fixes. The taskcluster tests in the puppet PR may miss this if it involves a # pyup: nodownload module.

If there’s a configuration problem, hopefully it’s clear by poking around the machine. If it’s pinned to an env, that may be part of it. I found a virtualenv that wouldn’t allow upgrading a package because someone installed the previous version as root, so cltbld couldn’t update. a chown -R cltbld /builds/scriptworker/lib fixed it.

There may be other failure cases; let’s update this list as we find them.

signing server¶

If we update the signing server dependencies, we probably need to restart all instances. Docs for restarting are [here](https://mana.mozilla.org/wiki/display/RelEng/Signing#Signing-(Re%29starting). The host list is here. The list of people who can ssh in is here (\$shortlist)

Success¶

When everything puppetizes successfully, we’re good!