PyYAML 4.2 Release Plan

[ingydotnet]

Synopsis

See the project planning page: https://github.com/yaml/pyyaml/projects/1

• Make release/4.2 with current master
• Make PR to revert Make pyyaml safe by default. #74 ( Reverting https://github.com/yaml/pyyaml/pull/74 #194 )
• PyYAML team fixes broken pyyaml-build system
  • Builds wheels with libyaml-2.1 linked in
• Fix any other 4.2 blocker issues
• Merge in successor to (Make pyyaml safe by default. #74, Rework the safety related API code #189) IF it reaches approval consensus
• Release 4.2 to PyPI
• Continue to work on successor if not part of 4.2
  • Bump version to 5.1 when merged

The PyYAML Release Situation

The most recent PyYAML, 3.12, was released Aug 2016. At that time, Kirill
turned over maintenance of PyYAML and LibYAML to @sigmavirus24 and @ingy .
Since then about 20 PRs have been applied to PyYAML and about 40 to LibYAML.

PyYAML has a release builder: https://github.com/yaml/pyyaml-build
It builds PyYAML wheels against specific versions combinations of
(Python, PyYAML, LibYAML).

This builder no longer works and it's complicated by the fact that the build
process for libyaml has been changed. The PyYAML team is working hard to fix it.

The 4.1 release attempt was rushed out because we knew that PyYAML 3.12 doesn't
work with Python 3.7 which went out this week. We had a fix for that in master,
and so we tried to get it out in time for 3.7.

We thought we had a Jenkins build system that would build the wheels as soon as
the sdist was uploaded. So we pushed the release only to find out minutes later
that this build system wasn't set up to build with libyaml. We were going to
have to use the pyyaml-build system.

After 48 hours of work on the windows/wheels system we decided to pull the plug
on 4.1. We didn't have wheels and we were getting reports of other things that
were wrong. We didn't have a sense that the build system was going to get fixed
soon, and we are all volunteers with limited time.

Soon after the release I learned about PR #74 and was completely surprised to
find that something this big went in without my seeing it. Looking back now I
remember that I had a lot going on in my life at that specific time.

#74 is a non backwards compatible change at the most basic level. It changes
how the dump and load functions behave.

The intent of the change is a good one:

• Currently PyYAML has the sugar-API: dump, load, safe_dump, safe_load
• PyYAML has had that API since version 3.1 (April 2006)
• load() is trivial to exploit on untrusted data
• Change load and dump to be aliases to safe_load and safe_dump
• Add alarming (danger_*) new functions for the old load and dump

But this change has contentions:

• It's non-back-compat and is going to affect a ton of existing code
• The name danger is misleading when used in completely safe ways
  • In addition, danger_dump is not known to be exploitable in any way
• PyYAML has known about this and had a safe_ solution in place from the start
  • People are just not feeling comfortable with the defaults

The change is important, worthy of a major release, but is not ready to
be part of PyYAML in its current form. A new PR, building from #74 and #189
should be worked on.

The Current Plan Forward

We need to get PyYAML released soon, if only for the Python 3.7 release. We
can't make any release at all until the build system works again. IOW, we
couldn't even re-release 3.12 right now.

The #74 API change is big and it is more important to get it right than to rush
it out. ie It load() may be a big can of gasoline, but nothing's on fire. ie
#74 doesn't "fix" anything. It just changes a default to something that's
always been safe and available.

There are 60 other changes that I'd like to tackle in the next release, while
at the same time taming a broken release process. My hope is that when we
figure this out, it will be easy to put out PyYAML releases on a regular basis.

We went from 3.12 to 4.x because this was a big release. It's big with or
without #74. I would like to see PyYAML 4.2 get out in the next few days.

If the successor to #74 / #189 is ready and approved by the time we are ready
to upload 4.2, it can go in.

If not, then I think it should be the focus of a 5.1 release. It's a big enough
change to trigger a major release. It should be in the first release of either
4.x or 5.x.

[asottile]

imo #74 is the most important change to keep in the 4.x release -- without it this is still a thing:

>>> yaml.danger_load("!!python/object/apply:os.system\nargs: ['echo I am a shell, wow']")
I am a shell, wow
0

I'd be pretty sad to see load() go back to its dangerous defaults.

[penguinolog]

Bump of major version number allows compatibility break, but it should be explained for everyone. And removal of release is worse, than just blacklisting release number as buggy (you can see openstack global requirements with a lot blacklisted versions).
P.S.
Today is bench of open non-destructive pull requests with bugfixes, which affect users and I expect, that some of them can be in the next release

[adamcharnock]

Thank you for working on this @ingydotnet. I really appreciate both the work going into sorting this out, and the clear communication above. I'm very pleased to hear that the project is moving forward again 👍

[cdent]

This is clearly a tricky situation but it seems like this plan can address it:

• Whatever the eventual outcome on Make pyyaml safe by default. #74 and Rework the safety related API code #189 it's clear that 4.1 was broken in several ways (including bad wheels and inconsistencies like PyYAML 4.1 changes "safe" in more ways than immediately obvious #187). Whether removal was the right thing (removal of 4.1 broke downstream gates #192) is moot at this point because it's already happened, some kind of plan needs to happen for moving forward.
• There's desire to have a working release for Python 3.7.
• There isn't agreement on how safe-by-default ought to work and the implementation in Make pyyaml safe by default. #74 broke people in weird ways (by changing what safe means). Getting agreement on how it should be done is important for the long-term health of PyYAML, otherwise, people will be presenting complaints and issues to a group of maintainers who never decided what right was, and it'll get confusing and stressful.
• If something was unsafe for several years, and it unsafe for a few more weeks, is that the end of the world?

So this plan seems okay. It addresses the need for a py3.7 release, avoids breaking people (other than #192) and gives time to figure out how to do safe-by-default well.

[ddanier]

Perhaps a bare Python 3.7 compatibility release could go into something like 3.13 and 4.2 could then include #74 and #189 and recent changes?

The missing compatibility with 3.7 currently breaks usage of PyYAML. This is definitely worse than not having #74 and #189 done properly I think.
Note: Homebrew just rolled out the Python 3.7 update. All packages using PyYAML are broken right now.

[asottile]

that would unblock python3.7 for me:

git checkout 3.12 -b 3.12.post1
sed -i 's/3\.12/3.12.post1/g' -- setup.py
virtualenv venv
venv/bin/pip install cython
venv/bin/python setup.py sdist

This produces a source distribution with working .c file that's installable in python3.7:

$ ls dist/
PyYAML-3.12.post1.tar.gz
$ venv/bin/pip uninstall cython -y >& /dev/null
$ venv/bin/pip install dist/PyYAML-3.12.post1.tar.gz
Processing ./dist/PyYAML-3.12.post1.tar.gz
Building wheels for collected packages: PyYAML
  Running setup.py bdist_wheel for PyYAML ... done
Successfully built PyYAML
Installing collected packages: PyYAML
Successfully installed PyYAML-3.12.post1
$ venv/bin/python --version
Python 3.7.0
$ venv/bin/python -c 'from yaml.cyaml import CLoader'
$

EDIT (so it's clear what the above does):

it applies this diff:

diff --git a/setup.py b/setup.py
index 9dc5e8d..0a1db91 100644
--- a/setup.py
+++ b/setup.py
@@ -1,6 +1,6 @@
 
 NAME = 'PyYAML'
-VERSION = '3.12'
+VERSION = '3.12.post1'
 DESCRIPTION = "YAML parser and emitter for Python"
 LONG_DESCRIPTION = """\

(just a version bump)

and then produces a new source distribution with freshly built .c file (due to new-enough cython being installed alongside the setup.py invocation)

[ingydotnet]

@ddanier @asottile your wish has (almost) come true...

@nitzmahone and I just released https://pypi.org/project/PyYAML/3.13b1/#files which is a rebuild of 3.12 with the latest cython. It includes 9 out of the 10 windows wheels we want to ship.

I expect the final 3.13 to be out in 24-48 hours. @nitzmahone is running tons of ansible tests to give us confidence.

We realized today that putting out a 3.13 PyYAML with a 0.1.7 LibYAML could happen a couple days sooner than a 4.2 with 0.2.2, so we've decided that having PyYAML work on Python 3.7 comes first.

This hasn't slowed down our work on 4.2, which we expect out possibly a week after 3.13. There's even a decent chance it can include a proper safe-load-by-default fix. @ddanier, remember that 4.2 is first about 60+ PRs to PyYAML and LibYAML since Aug 2016 (not just the #74 #189 #198 change).

[ddanier]

Thanks a lot!