Python 3.9 package is missing dependency on tzdata

[Christian Heimes]

Python 3.9 introduced the stdlib module "zoneinfo". The module either uses the operating sytem's tzdata database or the tzdata Python package. See https://www.python.org/dev/peps/pep-0615/ for more information.

Debian's and Ubuntu's python3.9 package is missing a dependency on tzdata. This breaks the zoneinfo module in a minimal container:

$ docker run -ti ubuntu:groovy /bin/bash
# apt update
# apt install -y python3.9
# dpkg -l python3.9
Desired=Unknown/Install/Remove/Purge/Hold
| Status=Not/Inst/Conf-files/Unpacked/halF-conf/Half-inst/trig-aWait/Trig-pend
|/ Err?=(none)/Reinst-required (Status,Err: uppercase=bad)
||/ Name Version Architecture Description
+++-==============-============-============-=============================================================
ii python3.9 3.9.0-5 amd64 Interactive high-level object-oriented language (version 3.9)
# python3.9 -c 'import zoneinfo; zoneinfo.ZoneInfo("CET")'
...
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/usr/lib/python3.9/zoneinfo/_common.py", line 24, in load_tzdata
    raise ZoneInfoNotFoundError(f"No time zone found with key {key}")
zoneinfo._common.ZoneInfoNotFoundError: 'No time zone found with key CET'

Anthony Sottile addressed the issue in his Deadsnakes PPA already, https://github.com/deadsnakes/python3.9/commit/d8216a1f585dac9db805d4e3c57f29535cf84cf8

[Christian Heimes (heimes)]

# export DEBIAN_FRONTEND="noninteractive"
# apt install -yq tzdata
# python3.9 -c 'import zoneinfo; print(zoneinfo.ZoneInfo("CET"))'
CET

[Matthias Klose (doko)]

tzdata is a required package. See Debian policy 2.5.

[Christian Heimes (heimes)]

Required or not, it's not part of the Ubuntu container image and therefore not available in minimal container setups. My initial comment contains a reproducer for the issue. It shows that zoneinfo module is broken in Ubuntu Groovy container.

[Paul G (p-ganssle)]

As I mentioned on the Arch Linux tracker, I think there are good reasons for declaring an explicit dependency on tzdata even if it the dependency should be satisfied transitively or implicitly: https://bugs.archlinux.org/task/68642?project=1&string=python

Presumably it doesn't hurt anything to explicitly add a dependency on this (since the reason for *not* doing so is that the package is already required), and it would prevent bugs like the one that Christian pointed out.

[Colin Watson (cjwatson)]

Matthias, your comment #2 is mistaken for two reasons:

 * In Ubuntu, tzdata is Priority: important, not Priority: required.
 * Only dependencies on Essential packages may be omitted; merely being Priority: required is not sufficient. See Debian policy 3.5.

[Dimitri John Ledkov (xnox)]

we just removed tzdata to be not required.

and it will make core snaps larger, because normally we exclude tzdata from minimal installs.

[Dimitri John Ledkov (xnox)]

Installing -full package will pull in all the things and make all the things work.

We absolutely have a wrath of people complaining about the size of tzdata and that it must not be in their containers / systems, even though they otherwise use python.

so would you rather not ship zoneinfo module, such that the import fails? and then move it to like python3-zoneinfo package and make regular std recommend it, but not depend on it.

That way classic server / desktop installs have it, but specifically docker containers will not, because they are required to ship without tzdata.

if i recall correctly, we only ship minimal docker containers with just the UTC timezone. Such that lookups of UTC should work.

At the moment, the way we ship python is a compromise between the wrath of people who deploy these things in production and demand to shrink the size of python and its deps. And the polar oposite of the typical desire to have all the things enabled / ship all the things.