Remove propagate_markers #7076
Remove propagate_markers #7076
Conversation
crates/uv-resolver/src/graph_ops.rs
Outdated
| /// marker), we re-queue the node and update all its children. This implicitly handles cycles, | ||
| /// whenever we re-reach a node through a cycle the marker we have is a more | ||
| /// specific marker/longer path, so we don't update the node and don't re-queue it. | ||
| pub(crate) fn marker_reachability<T>( |
There was a problem hiding this comment.
This function only moved.
There was a problem hiding this comment.
If possible I might suggest splitting these moves out into separate commits in the future. Otherwise it can be a little tricky to see which parts of the diff are "uninteresting moves" versus what is an interesting change.
(It can be hard to always do this though, especially if you do the move mid-refactor.)
konstin
force-pushed
the
konsti/propagate_markers
branch
from
64f60de
to
4b67ad1
Compare
crates/uv-resolver/src/graph_ops.rs
Outdated
| /// marker), we re-queue the node and update all its children. This implicitly handles cycles, | ||
| /// whenever we re-reach a node through a cycle the marker we have is a more | ||
| /// specific marker/longer path, so we don't update the node and don't re-queue it. | ||
| pub(crate) fn marker_reachability<T>( |
There was a problem hiding this comment.
If possible I might suggest splitting these moves out into separate commits in the future. Otherwise it can be a little tricky to see which parts of the diff are "uninteresting moves" versus what is an interesting change.
(It can be hard to always do this though, especially if you do the move mid-refactor.)
| @@ -201,7 +201,7 @@ impl std::fmt::Display for RequirementsTxtExport<'_> { | |||
| } | |||
| } | |||
|
|
|||
| if let Some(contents) = markers.contents() { | |||
| if let Some(contents) = self.reachability[&node_index].contents() { | |||
| // We use the markers of the base package: We know that each virtual extra package has an | ||
| // edge to the base package, so we know that base package markers are more general than the | ||
| // extra package markers (the extra package markers are a subset of the base package | ||
| // markers). |
There was a problem hiding this comment.
This seems like an interesting insight that might be worth copying over to the docs on PubGrubPackage. Specifically, is it a guaranteed invariant that for any package foo, if there is both a PubGrubPackage::Package and a PubGrubPackage::Marker for it, then the former is always a superset of the latter?
(I'd love to start collecting invariants in the docs for PubGrubPackage because I find that type somewhat difficult to reason about. But I don't have the full shape in my head yet to propose something better.)
There was a problem hiding this comment.
I have a hard time documenting this because "the base package is always selected when one of its virtual packages is selected" seems so fundamental to me. Do we have a good place to put overall resolver things? Because I think it would be valuable to write some of them up, we bet we come across some improvements to uv by this way.
There was a problem hiding this comment.
I'm not sure that was always even true, was it? The way virtual packages get added has shifted around a lot I think. Particularly in get_dependencies in the resolver. I don't think anything fundamental has changed in ~months, but I know @charliermarsh did a lot of work there.
I would start by adding these sorts of things to PubGrubPackage I think. It is a very key data type in the resolver and I think there are a lot of interesting invariants around it.
This is intrinsic to how we handle markers and to marker propagation. This property held true before and after, though the algorithm we arrived at it changed and i made it more explicit. |
konstin
force-pushed
the
konsti/propagate_markers
branch
from
4b67ad1
to
b7ff450
Compare
Follow-up to #6959 and #6961: Use the reachability computation instead of
propagate_markerseverywhere.With
marker_reachability, we have a function that computes for each node the markers under which it is (requirements.txt, no markers provided on installation) or can be (uv.lock, depending on the markers provided on installation) included in the installation. Put differently: If the marker computed bymarker_reachabilityis not fulfilled for the current platform, the package is never required on the current platform.We compute the markers for each package in the graph, this includes the virtual extra packages and the base packages. Since we know that each virtual extra package depends on its base package (
foo[bar]impliedfoo), we only retain the base package marker in therequirements.txtgraph.In #6959/#6961 we were only using it for pruning packages in
uv.lock, now we're also using it for the markers inrequirements.txt.I think this closes #4645, CC @bluss.