How to avoid circular dependencies when setting Properties?

Question

This is a design principle question for classes dealing with mathematical/physical equations where the user is allowed to set any parameter upon which the remaining are being calculated. In this example I would like to be able to let the frequency be set as well while avoiding circular dependencies.

For example:

from traits.api import HasTraits, Float, Property
from scipy.constants import c, h
class Photon(HasTraits):
    wavelength = Float # would like to do Property, but that would be circular?
    frequency = Property(depends_on = 'wavelength')
    energy = Property(depends_on = ['wavelength, frequency'])
    def _get_frequency(self):
        return c/self.wavelength
    def _get_energy(self):
        return h*self.frequency

I'm also aware of an update trigger timing problem here, because I don't know the sequence the updates will be triggered:

1. Wavelength is being changed
2. That triggers an updated of both dependent entities: frequency and energy
3. But energy needs frequency to be updated so that energy has the value fitting to the new wavelength!

(The answer to be accepted should also address this potential timing problem.)

So, what' the best design pattern to get around these inter-dependent problems? At the end I want the user to be able to update either wavelength or frequency and frequency/wavelength and energy shall be updated accordingly.

This kind of problems of course do arise in basically all classes that try to deal with equations.

Let the competition begin! ;)

Answer 1

Thanks to Adam Hughes and Warren Weckesser from the Enthought mailing list I realized what I was missing in my understanding. Properties do not really exist as an attribute. I now look at them as something like a 'virtual' attribute that completely depends on what the writer of the class does at the time a _getter or _setter is called.

So when I would like to be able to set wavelength AND frequency by the user, I only need to understand that frequency itself does not exist as an attribute and that instead at _setting time of the frequency I need to update the 'fundamental' attribute wavelength, so that the next time the frequency is required, it is calculated again with the new wavelength!

I also need to thank the user sr2222 who made me think about the missing caching. I realized that the dependencies I set up by using the keyword 'depends_on' are only required when using the 'cached_property' Trait. If the cost of calculation is not that high or it's not executed that often, the _getters and _setters take care of everything that one needs and one does not need to use the 'depends_on' keyword.

Here now the streamlined solution I was looking for, that allows me to set either wavelength or frequency without circular loops:

class Photon(HasTraits):
    wavelength = Float 
    frequency = Property
    energy = Property

    def _wavelength_default(self):
        return 1.0
    def _get_frequency(self):
        return c/self.wavelength
    def _set_frequency(self, freq):
        self.wavelength = c/freq
    def _get_energy(self):
        return h*self.frequency

One would use this class like this:

photon = Photon(wavelength = 1064)

or

photon = Photon(frequency = 300e6)

to set the initial values and to get the energy now, one just uses it directly:

print(photon.energy)

Please note that the _wavelength_default method takes care of the case when the user initializes the Photon instance without providing an initial value. Only for the first access of wavelength this method will be used to determine it. If I would not do this, the first access of frequency would result in a 1/0 calculation.

Answer 2

I would recommend to teach your application what can be derived from what. For example, a typical case is that you have a set of n variables, and any one of them can be derived from the rest. (You can model more complicated cases as well, of course, but I wouldn't do it until you actually run into such cases).

This can be modeled like this:

# variable_derivations is a dictionary: variable_id -> function
# each function produces this variable's value given all the other variables as kwargs
class SimpleDependency:
  _registry = {}
  def __init__(self, variable_derivations):
    unknown_variable_ids = variable_derivations.keys() - self._registry.keys():
      raise UnknownVariable(next(iter(unknown_variable_ids)))
    self.variable_derivations = variable_derivations

  def register_variable(self, variable, variable_id):
    if variable_id in self._registry:
      raise DuplicateVariable(variable_id)
    self._registry[variable_id] = variable

  def update(self, updated_variable_id, new_value):
    if updated_variable_id not in self.variable_ids:
      raise UnknownVariable(updated_variable_id)
    self._registry[updated_variable_id].assign(new_value)
    other_variable_ids = self.variable_ids.keys() - {updated_variable_id}
    for variable_id in other_variable_ids:
      function = self.variable_derivations[variable_id]
      arguments = {var_id : self._registry[var_id] for var_id in other_variable_ids}
      self._registry[variable_id].assign(function(**arguments))

class FloatVariable(numbers.Real):
  def __init__(self, variable_id, variable_value = 0):
    self.variable_id = variable_id
    self.value = variable_value
  def assign(self, value):
    self.value = value
  def __float__(self):
    return self.value

This is just a sketch, I didn't test or think through every possible issue.