PlanIndex.dependencyClosure does not follow setup dependencies(?)

[ezyang]

Hello @edsko,

Was it intentional for PlanIndex.dependencyClosure to not follow setup dependencies? It seems that it could go either way. Very subtle, and not documented!

When I change this function to follow setup dependencies, I find cycleThroughSetupDep5 starts failing. The reason is that when we construct an InstallPlan, we check to see if all of the versions of each transitive dependency tree are consistent. This process uses PlanIndex.dependencyClosure; and indeed if we don't chase setup dependencies we don't end up with inconsistent versions in our plans. I suppose it works out, but it is very finely balanced. To whit, I don't think we are able to detect if the solver produces an inconsistent set of versions for the setup dependency itself!

I encountered this while working on #3525. For now I'll hew to the old behavior, but I think this is a bit too tricky.

[edsko]

When we introduced setup dependencies, we used this policy: any function that wasn't following setup dependencies previously still didn't follow setup dependencies after their introduction. This seemed the most "obvious" way to migrate the existing code, with the option of modifying individual bits of code when found necessary. @dcoutts might have more insights here.

[ulysses4ever]

This looks like a question that was answered. Should we close it?

[Mikolaj]

Indeed, no followup questions in 6 years might indicate so. :) Closing.