[docs/convention] VariablePrimal: what should solvers return when there is no solution?

[dourouc05]

When working on Chuffed.jl, I wondered what I should return for VariablePrimal when the model is infeasible. Currently, this is not documented (https://github.com/jump-dev/MathOptInterface.jl/blob/master/src/attributes.jl#L1188-L1195).

Intuitively, two solutions make sense to me: either throw an appropriate exception (similarly to InvalidCallbackUsage or OptimizeInProgress) or return nothing.

[odow]

You're fine to return NaN or throw an error. People should only check VariablePrimal if PrimalStatus indicates a solution is present. This is purposefully undefined behavior because solvers do different things.

[dourouc05]

Would it be possible to state clearly in the docs that this is undefined behaviour? (Which is done for C and C++, for instance.) This could be useful for users when they encounter a mismatch between solvers and for solver implementers (mostly).

[odow]

Would it be possible to state clearly in the docs that this is undefined behaviour?

Sure. Make a PR?