Setting integrality / binary undefined in managed SCIP mode about scip.jl

[matbesancon]

Indeed this is maybe a pointless layer. I was thinking about it because the managed tests still required low-level manipulations as:

        # indicator constraint: z = 1 ==> 𝟙^T [x, y] <= 1.
        z = SCIP.add_variable(mscip)
        SCIP.@SC SCIP.SCIPchgVarType(mscip, SCIP.var(mscip, z), SCIP.SCIP_VARTYPE_BINARY, Ref{SCIP.SCIP_Bool}())
        SCIP.@SC SCIP.SCIPchgVarLb(mscip, SCIP.var(mscip, z), 0.0)
        SCIP.@SC SCIP.SCIPchgVarUb(mscip, SCIP.var(mscip, z), 1.0)

[rschwarz]

I kind of agree, in terms of consistency, but there are two special properties for these "variable" constraints (bounds and integral types):

• they are not considered constraints by SCIP (with their own SCIP_CONS), so they do not need reference counting for memory management (in the scope of ManagedSCIP).
• there is some not-so-nice setting of implicit bounds of [0, 1] for binary variables, together with storing and resetting older bounds. Would we have to handle this logic twice, at the ManagedSCIP and also at the MOI level?

[matbesancon]

Yes the idea was moving all parts requiring the Clang-generated API to managed, so that the dependency on the generated part is only in there (keeping it consistent with the other constraints).

Yes I saw the bounds, I think this could be handled once at the ManagedSCIP level, and only passed and not stored at MOI level

[rschwarz]

OK, I'm open to that kind of change in this case.

But I wonder whether it makes sense to require a layer between all calls from the MOI interface implementation to the Clang wrapper, in general. For example, a lot of result queries (in #115) are just one-liners.

I also expect power users to make some direct calls to (currently) unwrapped calls. It is already relatively convenient to do so, as SCIP.Optimizer and ManagedSCIP are automatically cconverted to Ptr{SCIP_}.