Latest revision as of 20:42, 23 September 2019

Limit Scriptlet Usage of core packages

Summary

Remove direct scriptlet calls from "core packages" (those that are used to build minimal container image). The packages can still affect changes during installation by placing files in the correct locations to trigger registered external programs.

Owner

Name: James Antill
Email: james@fedoraproject.org

Current status

Targeted release: Fedora 32
Last updated: 2019-09-23
Tracker bug: #1754666
Release notes tracker: #387

Detailed Description

Currently we know how to make an installable OS with packages that doesn't require the use of scriptlets, indeed rpm-ostree and others have already done this on a significantly bigger scale. So we plan to remove direct scriptlets from most (if not all) of the packages in the main fedora container image for Fedora 31. This means all four of: %pre/%post/%preun/%postun. After this change it'd be good to have some kind of temporary exception to be granted before those packages could add a scriptlet back (post F31 work).

Almost all of the hard work is already done, as rpm can react to files being dropped in specified places with known actions (Eg. In this way systemd components can create users or files). There are a few minor changes needed to packages to move from the old way of doing things (Eg. calling adduser) to doing the new thing. Note that while a program will still be run at installation time, those programs will be few and easily audited (as against the 666 slightly different ways of adding a user we currently have).

Benefit to Fedora

At the moment all of the packages with scriptlets (and anything installed with them) are basically opaque programs that happen to also carry files that are installed, from the point of view of all of the packaging tools (rpm, ostree, composer, etc.) . After this change the entire installation of the main container image will be declarative, and thus. understandable by all of those tools. This should make things faster, and allow new optimizations and features.

This should also make packagers life easier in the long term, as packaging becomes less unique.

Scope

Proposal owners:

James Antill
JamesAntill/ScriptletUsage
needs to get combine.d into the distribution, and then /etc/shells can use that.
Not so minor wrangling of package owners to tweak specfiles.

Other developers: N/A (not a System Wide Change)

Release engineering: #Releng issue number (a check of an impact with Release Engineering is needed)

Policies and guidelines: We should work toward only allowing new scriptlets to appear with policy exceptions, in any of the fixed packages. This needs to be done somewhat carefully, and post F31.

How To Test

All of the following should provide no output on a standard container:

* rpm -q -a --qf '%{preinprog}\n' | grep -v '(none)'
* rpm -q -a --qf '%{preunprog}\n' | grep -v '(none)'
* rpm -q -a --qf '%{postinprog}\n' | grep -v '(none)'
* rpm -q -a --qf '%{postunprog}\n' | grep -v '(none)'
* rpm -q -a --qf '%{pretransprog}\n' | grep -v '(none)'
* rpm -q -a --qf '%{posttransprog}\n' | grep -v '(none)'

User Experience

There should be no significant difference to users, initially, although package installation should be faster.

Dependencies

N/A (not a System Wide Change)

Contingency Plan

Until higher level packaging/installation tooling starts to depend on the scriptlets not being present it is always possible to just let those packages which still have scriptlets continue, if it's required.

Contingency mechanism: Just ship packages with scriptlets, as we always have done.
Contingency deadline: N/A
Blocks release? No

Documentation