Changed behavior of Preference.Any by pure python installation

[cmalinmayor]

ilpy/src/ilpy/solver_backends/__init__.py

Lines 30 to 40 in 6c35ace

	for modname, clsname in to_try:
	import_mod = f"ilpy.solver_backends.{modname}"
	try:
	mod = __import__(import_mod, fromlist=[clsname])
	cls = getattr(mod, clsname)
	backend = cls()
	assert isinstance(backend, SolverBackend)
	return backend
	except Exception as e: # pragma: no cover
	errors.append((f"{import_mod}::{clsname}", e))

Previously, if you put Preference.Any, and you didn't have a Gurobi license, we would try to create a gurobi solver and it would fail here. Gurobi when installed from pip/conda ships with a size-limited license (https://support.gurobi.com/hc/en-us/articles/360051597492-How-do-I-resolve-a-Model-too-large-for-size-limited-Gurobi-license-error). So, passing Preference.Any will always make a Gurobi solver, but later trying to solve any problem with more than 2000 parameters causes an error.

I mostly wanted to document this change and see what @tlambert03 thinks is the correct thing to do about it (if anything). Some options that I'm okay with:

• Accept the new behavior and tell users to specify SCIP if they don't have a gurobi license and want to solve large-ish problems.
• Deprecate "Preference.Any" and force users to be explicit, since now it is harder to infer which solver to use.
• Check if the license is size limited and fall back to SCIP if so, to try and emulate the previous behavior. (I assume this is possible but I didn't look into how)

[tlambert03]

aha! yeah that's no good. I think that, behaviorally, we should return to the previous state of things, where gurobi is only used if a license is available. I don't imagine it will be hard to programmatically determine that at runtime. So yeah, your option 3 is what I would pick. We might also consider a new preference Preference.GurobiUnlicensed that would be tried after Preference.Gurobi, Preference.Scip, and would still let someone use it with the limited version.

[cmalinmayor]

Agreed on option 3! If you have time, it would be great if you would take a stab at implementation. Otherwise, I'll get to it when I get to it.