Hmm. I don't think we should allow get to return incorrect results or say that set! values can be ignored. These should be errors.

I think that ideally:

• It should always be possible to get any attribute in a 'safe' way, e.g. by having a checkedget/safeget (or maybe this should be called get) that first checks canget and then returns the result. This should probably be the default behavior.
• There should be an unsafeget version (currently called get) for at least some properties that foregoes the canget check for efficiency reasons, (e.g when you want to iterate over a bunch of VariableIndexes) and you've already checked canget for ::Type{VariableIndex}.

This is starting to sound like the @inbounds machinery. Unfortunately, that's hard to replicate, but maybe some kind of simple @unsafe macro solution to redirect to the unsafeget version would be nice.

@tkoolen is there really a need for an unsafe version? If you're iterating over a bunch of VariableIndexes then you can use the ::Vector{VariableIndex} version of get instead.

What would be done in unsafeget if canget is false ? The solvers returns an incorrect value and the error is thrown in safeget which is implemented in MOI like this ?

function MOI.safeget(...)
    if MOI.canget(...)
        MOI.get(...)
    else
        error(...)
    end
end

@mlubin, I thought that there was a need for that given your earlier comment, JuliaOpt/MathOptInterfaceBridges.jl#87 (comment). If that's not the case, what I proposed is probably unnecessary.

@blegat, yep.

This has been updated to reflect that the call should throw an error.

Closing in favour of #228