Custom attributes not queried correctly with caching optimizer

[odow]

See https://stackoverflow.com/questions/61922425/how-can-i-determine-whether-a-jump-model-solved-by-gurobi-is-a-mip

Issue is in functions like this:
https://github.com/JuliaOpt/MathOptInterface.jl/blob/3f872f9edb50b1a97cb65011430a0160fd56111e/src/Utilities/cachingoptimizer.jl#L539-L557
where we need a branch to see if it is possible to query directly from the optimizer.

[blegat]

One option would be to make ModelAttribute("IsMIP") be an optimizer attribute. The reasoning of the caching optimizer code is: If it's a model attribute then as the cache has the full model, it should be able to answer to get. Here the cache would be able to answer to IsMIP, it can just check whether there are integer/binary variables.
So another option would be to make this a solver-independent attribute. In general, I think solvers should not define custom model attributes that should be get but never set. If it's a custom attribute that can be set and then get (like callbacks) then it's ok for it to be model attributes as the cache will also have the attribute set so it can answer to get.

[odow]

One option would be to make ModelAttribute("IsMIP") be an optimizer attribute.

But there are also variable and constraint attributes.

In general, I think solvers should not define custom model attributes that should be get but never set.

I don't think this is a workable solution. We have a counter-example with Gurobi, and there are bound to be other examples.

[blegat]

We could have a function similar to MOI.is_set_by_optimize like MOI.is_set_by_optimizer. Here it seems the attribute is set by the optimizer even when TerminationStatus is OPTIMIZE_NOT_CALLED so returning true for is_set_by_optimize would not work, correct ?

[odow]

is_set_by_optimizer seems sensible. Or something like is_optimizer_specific.

[odow]

Conclusion from the call is that Gurobi should wrap the attribute in
https://github.com/JuliaOpt/MathOptInterface.jl/blob/1315e317da1d59c2cb172145a8699b3e0a6135b8/src/Utilities/cachingoptimizer.jl#L657-L659

[odow]

Actually, we probably want is_attribute_from_optimizer(attr) so that users don't have to do this.

[blegat]

Actually, we probably want is_attribute_from_optimizer(attr) so that users don't have to do this.

Sounds good

[odow]

I think this could be better handled via JuMP:

function JuMP.get_optimizer_attribute(
    model::Model,
    attr::MOI.AbstractModelAttribute,
)
    return MOI.get(backend(model), MOI.Utilities.AttributeFromOptimizer(attr))
end

model = Model(Gurobi.Optimizer)
MOI.Utilities.attach_optimizer(model)
get_optimizer_attribute(model, Gurobi.ModelAttribute("IsMIP"))

At the MOI level, it's sufficient to wrap attributes in AttributeFromOptimizer.

[odow]

This works. Needs docs and tests

function JuMP.get_optimizer_attribute(
    model::Model, 
    attr::MOI.AbstractModelAttribute,
)
    b = backend(model)
    if mode(model) == DIRECT
        return MOI.get(b, attr)
    end
    return MOI.get(b, MOI.Utilities.AttributeFromOptimizer(attr))
end

function JuMP.get_optimizer_attribute(
    model::Model,
    attr::MOI.AbstractConstraintAttribute,
    c::ConstraintRef,
)
    b = backend(model)
    if mode(model) == DIRECT
        return MOI.get(b, attr, index(c))
    else
        return MOI.get(b, MOI.Utilities.AttributeFromOptimizer(attr), index(c))
    end
end

function JuMP.get_optimizer_attribute(
    model::Model,
    attr::MOI.AbstractVariableAttribute,
    x::VariableRef,
)
    b = backend(model)
    if mode(model) == DIRECT
        return MOI.get(b, attr, index(x))
    else
        return MOI.get(b, MOI.Utilities.AttributeFromOptimizer(attr), index(x))
    end
end

model = Model(Gurobi.Optimizer)
@variable(model, x >= 1.5)
@constraint(model, c, 2x + 1 == 0)
MOI.Utilities.attach_optimizer(model)
get_optimizer_attribute(model, Gurobi.ModelAttribute("IsMIP")) == 0
get_optimizer_attribute(model, Gurobi.VariableAttribute("LB"), x) == 1.5
get_optimizer_attribute(model, Gurobi.ConstraintAttribute("RHS"), c) == -1.0

[joaquimg]

How much of that extends to other solvers?
Or is that Gurobi's interface?

Also, what about cases where you want to query attributes from a pair of variables (in a quadratic) and a pair variable-constraint (coefficients for instance).

I had a use case in DiffOpt.jl where we want to query derivatives from all problem data.

Maybe we can have a more generic method with some "args..." and pass them through a convert_to_index_if_needed function?

[odow]

Closing in favor of jump-dev/Gurobi.jl#425