In the first scenario, preferences-dependent-forking, we show how the resolution can be different the second time because we don't fork anymore because the preferences from the lockfile make us skip the fork point. This scenario is inspired by the same happening in the transformer repo, where a single universal resolution leads to a more desirable lock without duplicate versions.
Consider a fresh run without preferences:
We start with cleaver 2
We fork
We reject cleaver 2
We find a cleaver 1 solution in fork 1 with foo 2 with bar 1
We find a cleaver 1 solution in fork 2 with foo 1 with bar 2
We write cleaver 1, foo 1, foo 2, bar 1 and bar 2 to the lockfile
In a subsequent run, we read the preference cleaver 1 from the lockfile (the preferences for foo and bar don't matter):
We start with cleaver 1
We're in universal mode, we resolve foo 1 and bar 1
We write cleaver 1 and bar 1 to the lockfile
We call a resolution that's different between first and second run unstable.
Design sketch:
root -> cleaver, foo, bar
# Cause a fork, then forget that version.
cleaver 2 -> unrelated-dep==1; fork==1
cleaver 2 -> unrelated-dep==2; fork==2
cleaver 2 -> reject-cleaver-2
# Allow diverging versions when forking, but force foo 1, bar 1 in universal mode without forking.
cleaver 1 -> foo==1; fork==1
cleaver 1 -> bar==1; fork==2
Let's try this out:
[project]
name = "dummy"
version = "0.1.0"
requires-python = ">=3.12"
dependencies = [
"preferences-dependent-forking-5c5f76e6"
]
[tool.uv]
index-url = "http://127.0.0.1:3141/"
Cleanup and first, fresh run without preferences, on linux:
$ rm -f uv.lock
$ uv venv -q
$ cargo run -q -- sync --preview
Resolved 6 packages in 647ms
Built dummy @ file:///home/konsti/projects/uv/debug/dummy
Built preferences-dependent-forking-5c5f76e6==0.0.0
Prepared 2 packages in 457ms
Installed 5 packages in 1ms
+ dummy==0.1.0 (from file:///home/konsti/projects/uv/debug/dummy)
+ preferences-dependent-forking-5c5f76e6==0.0.0
+ preferences-dependent-forking-bar-5c5f76e6==1.0.0
+ preferences-dependent-forking-cleaver-5c5f76e6==1.0.0
+ preferences-dependent-forking-foo-5c5f76e6==2.0.0
Second run, using the preferences from the lockfile:
$ cargo run -q -- sync --preview
Resolved 5 packages in 113ms
Uninstalled 1 package in 0.57ms
Installed 1 package in 1ms
- preferences-dependent-forking-foo-5c5f76e6==2.0.0
+ preferences-dependent-forking-foo-5c5f76e6==1.0.0
The logs confirm that in the first run with fork, while in the second we don't.
In the second scenario preferences-dependent-forking-conflicting, we add foo 1 -> bar==2 so that the resolving cleaver 1 fails without forking:
[project]
name = "dummy"
version = "0.1.0"
requires-python = ">=3.12"
dependencies = [
"preferences-dependent-forking-conflicting-5f0792c1"
]
[tool.uv]
index-url = "http://127.0.0.1:3141/"
Cleanup and first, fresh run, on linux:
$ rm -f uv.lock
$ uv venv -q
$ cargo run -q -- sync --preview
Resolved 7 packages in 166ms
Built dummy @ file:///home/konsti/projects/uv/debug/dummy
Prepared 1 package in 489ms
Installed 5 packages in 1ms
+ dummy==0.1.0 (from file:///home/konsti/projects/uv/debug/dummy)
+ preferences-dependent-forking-conflicting-5f0792c1==0.0.0
+ preferences-dependent-forking-conflicting-bar-5f0792c1==2.0.0
+ preferences-dependent-forking-conflicting-cleaver-5f0792c1==1.0.0
+ preferences-dependent-forking-conflicting-foo-5f0792c1==2.0.0
In the second run, we're using the preferences from the lockfile. We reject preference cleaver 1 in universal mode, then move to cleaver 2, fork, and fail in one of the forks by the cleaver 2 rejecting package:
$ cargo run -q -- sync --preview
× No solution found when resolving dependencies for split (sys_platform != 'linux'):
╰─▶ Because preferences-dependent-forking-conflicting-foo-5f0792c1==1.0.0
depends on preferences-dependent-forking-conflicting-bar-5f0792c1==2 and
preferences-dependent-forking-conflicting-cleaver-5f0792c1==1.0.0 depends on
preferences-dependent-forking-conflicting-foo-5f0792c1{sys_platform == 'linux'}==1, we can
conclude that preferences-dependent-forking-conflicting-cleaver-5f0792c1==1.0.0 depends on
preferences-dependent-forking-conflicting-bar-5f0792c1==2.
And because preferences-dependent-forking-conflicting-cleaver-5f0792c1==1.0.0 depends on
preferences-dependent-forking-conflicting-bar-5f0792c1{sys_platform != 'linux'}==1 and only the following
versions of preferences-dependent-forking-conflicting-cleaver-5f0792c1 are available:
preferences-dependent-forking-conflicting-cleaver-5f0792c1==1.0.0
preferences-dependent-forking-conflicting-cleaver-5f0792c1==2.0.0
we can conclude that preferences-dependent-forking-conflicting-cleaver-5f0792c1<2.0.0 cannot be used. (1)
Because preferences-dependent-forking-conflicting-reject-cleaver-2-5f0792c1==1.0.0
depends on preferences-dependent-forking-conflicting-unrelated-dep-5f0792c1==3 and only
preferences-dependent-forking-conflicting-reject-cleaver-2-5f0792c1==1.0.0 is available, we can
conclude that all versions of preferences-dependent-forking-conflicting-reject-cleaver-2-5f0792c1 and
preferences-dependent-forking-conflicting-unrelated-dep-5f0792c1{sys_platform != 'linux'}==2.0.0 are
incompatible.
And because preferences-dependent-forking-conflicting-cleaver-5f0792c1==2.0.0 depends on
preferences-dependent-forking-conflicting-unrelated-dep-5f0792c1{sys_platform != 'linux'}==2
and preferences-dependent-forking-conflicting-reject-cleaver-2-5f0792c1, we can conclude that
preferences-dependent-forking-conflicting-cleaver-5f0792c1==2.0.0 cannot be used.
And because we know from (1) that preferences-dependent-forking-conflicting-cleaver-5f0792c1<2.0.0 cannot
be used, we can conclude that all versions of preferences-dependent-forking-conflicting-cleaver-5f0792c1
cannot be used.
And because preferences-dependent-forking-conflicting-5f0792c1==0.0.0 depends on
preferences-dependent-forking-conflicting-cleaver-5f0792c1, we can conclude that
preferences-dependent-forking-conflicting-5f0792c1==0.0.0 cannot be used.
And because only preferences-dependent-forking-conflicting-5f0792c1==0.0.0 is available and dummy==0.1.0
depends on preferences-dependent-forking-conflicting-5f0792c1, we can conclude that dummy==0.1.0 cannot
be used.
And because only dummy==0.1.0 is available and you require dummy, we can conclude that the requirements
are unsatisfiable.
The logs confirm that we fork in both runs.
I don't know whether this scenario should be tagged satisfiable yet or not.
In the first scenario,
preferences-dependent-forking
, we show how the resolution can be different the second time because we don't fork anymore because the preferences from the lockfile make us skip the fork point. This scenario is inspired by the same happening in thetransformer
repo, where a single universal resolution leads to a more desirable lock without duplicate versions.Consider a fresh run without preferences:
In a subsequent run, we read the preference cleaver 1 from the lockfile (the preferences for foo and bar don't matter):
We call a resolution that's different between first and second run unstable.
Design sketch:
Let's try this out:
Cleanup and first, fresh run without preferences, on linux:
Second run, using the preferences from the lockfile:
The logs confirm that in the first run with fork, while in the second we don't.
In the second scenario
preferences-dependent-forking-conflicting
, we addfoo 1 -> bar==2
so that the resolving cleaver 1 fails without forking:Cleanup and first, fresh run, on linux:
In the second run, we're using the preferences from the lockfile. We reject preference cleaver 1 in universal mode, then move to cleaver 2, fork, and fail in one of the forks by the cleaver 2 rejecting package:
The logs confirm that we fork in both runs.
I don't know whether this scenario should be tagged
satisfiable
yet or not.