Consider the following models:
class Operand(models.Model):
value_1 = models.CharField(maxlength=200)
value_2 = models.CharField(maxlength=200)
class Operation(models.Model):
arg = models.ForeignKey(Operand)
Now let's suppose we want to do some work on Operands:
>>> common_arg = Operand(value_1="val1", value_2="val2")
>>> common_arg.save()
>>> op1 = Operation(arg = common_arg)
>>> op2 = Operation(arg = common_arg)
>>> op1.save()
>>> op2.save()
>>> operations = Operation.objects.all().select_related()
>>> operations
[<Operation: Operation object>, <Operation: Operation object>]
>>> operations[0].arg.value_1 = "new_val1"
>>> operations[0].arg.save()
>>> operations[1].arg.value_2 = "new_val2"
>>> operations[1].arg.save()
>>> modified_arg = Operand.objects.get(id=1)
# OOPS we lost one change here !!
>>> modified_arg.value_1
u'val1'
>>> modified_arg.value_2
u'new_val2'
# restore the initial state
>>> modified_arg.value_2 = "val2"
>>> modified_arg.save()
# this is not limited to select_related but pervasive with FKs
>>> work_to_do = [(op, op.arg) for op in Operation.objects.all()]
>>> work_to_do[0][1].value_1 = "new_val1"
>>> work_to_do[0][1].value_1
'new_val1'
# oops, operation 2 will use the outdated argument
>>> work_to_do[1][1].value_1
u'val1'
#17 could provide a solution for this issue: caching instances in memory so get() and friends can return the instance we already have instead of creating a new, distinct one.
The issue is especially easy to encounter with select_related because all related fields will be preloaded in distinct instances in memory. Walking Operation.objects.all() and
modifying the Operands from there would load an new instance from the DB for each operation and work as expected in most cases.
