How to store arbitrary name/value key pairs in a Django model?

Question

I have a fixed data model that has a lot of data fields.

class Widget(Models.model):
    widget_owner = models.ForeignKey(auth.User)
    val1 = models.CharField()
    val2 = models.CharField()
    ...
    val568 = ...

I want to cram even more data into this Widget by letting my users specify custom data fields. What's a sane way to do this? Is storing name/value pairs where the user can specify additional "Widget fields" a good idea? My pseudo thoughts are below:

data_types = ('free_text', 'date', 'integer', 'price')
class CustomWidgetField(models.Model)
  owner = ForeignKey(auth.User)
  field_title = models.CharField(auth.User)
  field_value_type = models.CharField(choices = data_types)

class CustomWidgetValue(models.Model)
  field_type = ForeignKey(CustomWidgetField)
  widget = ForeignKey(Widget)
  value = models.TextField()

So I want to let each user build a new type of data field that will apply to all of their widgets and then specify values for each custom field in each widget. I will probably have to do filtering/searching on these custom fields just as I would on a native field (which I assume will be much slower than operating on native fields.) But the scale is to have a few dozen custom fields per Widget and each User will only have a few thousand Widgets in their inventory. I can also probably batch most of the searching/filtering on the custom fields into a backend script (maybe.)

Answer 1

Consider representing all custom properties with serialized dict. I used this in a recent project and it worked really well.

 class Widget(models.Model):
      owner = models.ForeignKey(auth.User)
      props = models.TextField(blank=True) # serialized custom data

      @property
      def props_dict(self):
          return simplejson.loads(self.props)

 class UserProfile(models.Model)
      user = models.ForeignKey(auth.User)
      widget_fields = models.TextField(blank=True) # serialized schema declaration

Answer 2

It looks like you've reinvented the triple store. I think it's a common thing, as we follow the idea of database flexibility to its natural conclusion. Triple stores tend to be fairly inefficient in relational database systems, but there are systems designed specifically for them.

http://en.wikipedia.org/wiki/Triplestore

At the scales you're talking about, your performance is likely to be acceptable, but they don't generally scale well without a specialized DB.