`set_start_value` silently fails for optimizers which do not support it

[araujoms]

using Hypatia, JuMP

model = Model(Hypatia.Optimizer)
@variable(model, x)
set_start_value(x, 3)
@constraint(model, 5x-2 ≥ 0)
@objective(model, Min, x)
optimize!(model)

The line set_start_value(x, 3) is silently ignored, it has no effect on the solver. Same thing happens with Clarabel.

[odow]

This is intended behavior

MathOptInterface.jl/src/Utilities/copy.jl

Line 74 in f6f7974

if attr == MOI.VariableName() || attr == MOI.VariablePrimalStart()

Start values are a hint. We long ago decided not to error. You could reasonably come to a different conclusion, but we’re not making breaking changes to MOI at this point.

[araujoms]

You could show a warning.

[odow]

We could, but I’m not very keen. We’d just be cluttering up the screen. And this isn’t something the user can do something about. And it also doesn’t change the solution etc. This is a really an issue that should be opened in the relevant solver wrappers. It’s not something MOI should deal with. Also, the skipping is only if wrappers use the default copying in MOI. If they write their own copy_to they can choose what to do.

[araujoms]

The user can do something about it: stop providing the useless starting value. A student of mine just wasted a lot of time trying to warm-start Hypatia as she believed it was supported due to the absence of errors or warnings.

[blegat]

The line is here

MathOptInterface.jl/src/Utilities/copy.jl

Line 75 in f6f7974

continue # Skipping names and start values is okay.

The reasoning is : if you write a solver independent model with starting values and then run it through many solvers, you don't want it to error on the solvers that don't support them.
Now I agree it already took me a lot of debugging time that a clear error with stacktrace would have helped.
Not sure what we can do without a breaking change though.

Ideally, I think we should have errored or at least warned and defined an attribute that, when set, would silent these errors. The default copy-to could read this attribute to see whether it should error or not
We could define this attribute and make the silent mode be default to be non breaking.
The user would set it to error or warning to help debugging.

[araujoms]

The scenario where I wrote a solver independent model and I'm running it through many solvers it's where I want the error the most; otherwise to find out whether the starting values are being used I need to go through documentation or do numerical experiments.

I don't think an @warn maxlog = 1 would be breaking.

[odow]

This wouldn't help Hypatia because Hypatia doesn't use MOI.Utilities.default_copy_to.

https://github.com/jump-dev/Hypatia.jl/blob/a9d0a28578c290635d5589aed2916d4d70acb319/src/MathOptInterface/wrapper.jl#L111-L116

The appropriate next step for this issue is to open an issue in the relevant solver wrappers.

I don't want to change anything in MOI.

[odow]

Closing because this is intended behaviour.

I think we all agree that the correct behaviour is debatable, so I've added this as a point to #2180.

But the choice here is in a helper utility method. If solvers want to use @warn they're welcome to. But that's not something we can or should fix in MOI v1.