UniversalFallback should throw key error rather than return canget == false if attribute partially set

[mlubin]

https://github.com/JuliaOpt/MathOptInterface.jl/blob/7164fdc573f3f92acb93de1bf18a3d3bd5e315c7/src/Utilities/universalfallback.jl#L54-L57

The check on the length of the dictionary means that:

m = Model()
@variable(m, x, start = 1)
@variable(m, y)
JuMP.startvalue(x)

gives

 Attribute MathOptInterface.VariablePrimalStart() not accessible
  Stacktrace:
   [1] get(::MathOptInterface.Utilities.CachingOptimizer, ::MathOptInterface.VariablePrimalStart, ::MathOptInterface.VariableIndex) at /home/mlubin/.julia/v0.6/MathOptInterface/src/Utilities/cachingoptimizer.jl:380
   [2] get(::JuMP.Model, ::MathOptInterface.VariablePrimalStart, ::JuMP.Variable) at /home/mlubin/.julia/v0.6/JuMP/src/JuMP.jl:379
   [3] startvalue(::JuMP.Variable) at /home/mlubin/.julia/v0.6/JuMP/src/variables.jl:379

The attribute values should be available immediately after you set them, and there should be a key error in the case that you ask for a value that's not defined.

[blegat]

In the doc, we have say that canget returns true if it is

applied to every variable

https://github.com/JuliaOpt/MathOptInterface.jl/blob/9f75cea0bc305d0628827b29427a329809b0f33d/src/attributes.jl#L141-L143

[mlubin]

It's evident to me that this is a bad design. Attribute MathOptInterface.VariablePrimalStart() not accessible is an incorrect and misleading error message in this situation. If an attribute is partially set, the model is still responsible for that attribute, and get should be called. The model may choose to return an error if it can't access the attribute, and that's ok. We might need to brainstorm a bit on how to resolve this situation and #295 now that we have a bit more experience working with MOI.

[blegat]

When MOI.get is called on CachingOptimizer, and canget is false for both the model and the optimizer, it could try to still call get on one of them as you suggest. It should try get on the one that supports this even it is not set for all variables. supports is probably the solution then.

If both canget return false, CachingOptimizer calls support on the optimizer, if it returns true it calls get on it. Otherwise it calls get on the model (the model supports everything because of the UniversalFallback).