VariableConstraint

[blegat]

The way variables are represented internally by MOIU.GenericModel is encapsulated in the following 4 lines

MathOptInterface.jl/src/Utilities/model.jl

Lines 993 to 1001 in a1d3ed9

	num_variables_created::Int64
	# If nothing, no variable has been deleted so the indices of the
	# variables are VI.(1:num_variables_created)
	variable_indices::Union{Nothing,Set{VI}}
	# Union of flags of `S` such that a `SingleVariable`-in-`S`
	# constraint was added to the model and not deleted yet.
	single_variable_mask::Vector{UInt8}
	# Bounds set by `SingleVariable`-in-`S`:
	variable_bounds::Box{T}

We could factor them out into a ModelLike supporting only these SingleVariable constraints.
This would allow MOI wrapper that need this to just have this as a field as well and not have to redo all the work.
That would apply to all solvers currently having the VariableInfo struct.
The issue is that this is currently hardcoded for a few SingleVariable constraints of interest.

EDIT: just notice that this is close to what #1468 does

[odow]

Not in favor of this.

I think #1468 is sufficient in the near-term. We can add functionality to simplify solver interfaces post 1.0. (We keep having scope-creep on our 0.10 release, so I'd rather focus on that.) I also don't think it's breaking to modify internal fields of Model later on.

[blegat]

It is breaking as solvers access the lower and upper bound vectors directly:
https://github.com/jump-dev/Clp.jl/pull/114/files#diff-18c6af231f3facc5ba10dd45231da349efda22ecd8d68a4661dee8027bc413a7R310-R311

[odow]

Ah. I guess #1468 is almost exactly what you're after. It just needs a bit more logic moved from model to box.jl.

I might make a separate type for the MatrixOfConstraint use-case. it can re-use some of the logic, but having Box appear in two places is what's confusing.

[blegat]

As discussed in #1468 (comment), I don't see what's confusing. The box in itself is a well defined object. It's only constructed differently in MatrixOfConstraints and AbstractModel but it would make sense to reuse the same structure. In fact, the Box struct could also be an AbstractVectorSet which is the cartesian product of Interval's. This is a request we sometimes get, e.g. from COSMO (although MatrixOfConstraints might resolve COSMO's use case).

So it seems intuitive to me that we have a Box object which is a hyperrectangle (and might or might not be a AbstractVectorSet now or in the future) and a ModelLike object which stores SingleVariable constraints internally as a big VectorOfVariable-in-Box constraint.

[odow]

Closed by #1468

Solvers can use SingleVariableConstraints and handle supports_constraint separately.