Define supports for is_set_by_optimize attributes

[odow]

supports is currently restricted to copyable attributes. But it would also be useful to extend the definition to optimizer attributes so you can know whether to query the attribute without risking an error.

We've discussed this before, but I can't find the issue. I think it may have been a comment on a pull request.

Related: jump-dev/JuMP.jl#2704 (comment)

[odow]

I'm not convinced by this. I see it as too breaking because every solver will need to be updated.

[odow]

@blegat this is our remaining blocker.

[blegat]

Currently, we could require supports to be implemented and start testing it but not use it in CachingOptimizer yet. When we release v0.11 or v1.0, we start using it in CachingOptimizer in something like:

if has_fallback(attr) && !MOI.supports(model.optimizer, attr)
    return get_fallback(model, attr, ...)
else
    return MOI.get(model.optimizer, attr, ...)
end

For attributes such that has_fallback is false, it won't make any changes. So unless we test MOI.supports in the tests, the solvers won't notice the difference.

[odow]

I think the options are:

1. Do nothing. Release MOI v1.0 which is non-breaking compared with 0.10
2. Release a breaking v0.11 change with tests and updated code
3. Release a breaking v1.0 change with tests and updated code

I guess it's a pretty trivial update for authors upgrading to MOI 1.0 anyway; they just have to add MOI.supports(::Optimizer, ::MOI.VariablePrimal, ::MOI.VariableIndex) etc.

It's a bit annoying because there are other attributes like ConstraintFunction and ConstraintSet which are neither copyable or set by optimize.

[blegat]

For ConstraintFunction, it could be confusing since some solvers could support setting it but not getting it.

[odow]

Monthly call says: Don't add it, but document the reasons why, and the work-arounds.

Edit: the larger discussion raised these points:

• We get the motivation for adding this
• Adding it will cause large breaking changes
• We don't think the upside of adding outweighs the downside of the breaking changes. Perhaps we could revisit in MOI 2.0.
• PowerSimulations use try-catch to check what they can and cannot query. They could use has_duals(model), or cache the results.
• MOI uses a try-catch in the caching optimizer. To date, no one has complained about performance. If they did, we could cache the result.