Constraint's equality attribute sometimes fails to properly identify equality constraints

[bknueven]

Summary

Occasionally constraint's new equality attribute will fail in ways that are difficult to reproduce.

Steps to reproduce the issue

This happened while I was running the sequential decomposition tool, but I can't seem to create a small example which reproduces the issue on my system. Instead I've included some hopefully useful output from the debugger:

ipdb> type(con)
<class 'pyomo.core.base.constraint._GeneralConstraintData'>
ipdb> type(con._expr)
<class 'pyomo.core.expr.logical_expr.RangedExpression'>
ipdb> con.pprint()
{Member of flow_mass_phase_comp_equality} : Size=2, Index=fs.feed._flow_mass_phase_comp_outlet_ref_index, Active=True
    Key                 : Lower : Body                                                                                                                                       : Upper : Active
    (0.0, 'Liq', 'H2O') :   0.0 : fs.feed.properties[0.0].flow_mass_phase_comp[Liq,H2O] - fs.PrimaryPumps[1].control_volume.properties_in[0.0].flow_mass_phase_comp[Liq,H2O] :   0.0 :   True
ipdb> id(con._expr.arg(2))
140324945312400
ipdb> id(con._expr.arg(0))
140324945312432
ipdb> con._expr.arg(0) == con._expr.arg(2)
True
ipdb> type(con._expr.arg(0))
<class 'float'>
ipdb> type(con._expr.arg(2))
<class 'float'>
ipdb> con._expr.arg(0) is con._expr.arg(2)
False
ipdb> con.equality
False
ipdb> 

Information on your system

Pyomo version: 6.1.1
Python version: 3.7.9
Operating system: Ubuntu 20.04.2
How Pyomo was installed (PyPI, conda, source): source
Solver (if applicable): N/A

Additional information

I believe the issue is with the is check on bounds in the equality attribute on line 416:

pyomo/pyomo/core/base/constraint.py

Lines 408 to 418 in 3cf310a

	@property
	def equality(self):
	"""A boolean indicating whether this is an equality constraint."""
	if self._expr.__class__ is logical_expr.EqualityExpression:
	return True
	elif self._expr.__class__ is logical_expr.RangedExpression:
	# TODO: this is a very restrictive form of structural equality.
	lb = self._expr.arg(0)
	if lb is not None and lb is self._expr.arg(2):
	return True
	return False

For RangedExpressions with floating point bounds, CPython does not guarantee that the same valued float objects will occupy the same location in memory (as shown by the debugging output above). However, it is usually the case that identical floating point values do occupy the same location in memory, making this difficult to reproduce. There was some discussion around resolving this issue in Python itself: https://discuss.python.org/t/demoting-the-is-operator-to-avoid-an-identity-crisis/86, but the CPython team went a different route: https://bugs.python.org/issue34850.

[jsiirola]

I was thinking about this when implementing it. The question in my mind is if equality should be "structurally" or "effectively". That is, consider:

model.x = Var()
model.l = Param(initialize=0, mutable=True)
model.u = Param(initialize=0, mutable=True)
model.c = Constraint(expr=inequality(model.l, model.x, model.u)

Should model.c.equality be True or False? If the answer is True, would it be weird if the user changed model.u to be 1 and model.c.equality suddenly became False?

One alternative could be to only use is when _lower / _upper are non-constant Pyomo objects, and fall back on equality for native times. I think I like this compromise.

[bknueven]

I agree that is should be used for non-constant Pyomo objects for exactly your case. So your compromise makes good sense to me.

Note that the original constraint above was created by the network.expand_arcs transformation as var1 == var2. So in this instance the expression system created this issue by unluckily grabbing two separate instances of 0.. I might not have noticed except the sequential decomposition tool checks the equality attribute on the expanded arc constraints.

[jsiirola]

@bknueven, while #2097 resolves this specific behavior, I am not at all clear how you are getting a RangedExpression in your constraint to begin with. Can you share your model / driver (off-issue is fine)