mirror of
https://github.com/encode/django-rest-framework.git
synced 2025-08-04 12:30:11 +03:00
Add nested serialiser example
This commit is contained in:
parent
b06c8f29e3
commit
5d40079924
|
@ -685,8 +685,9 @@ the coordinate pair:
|
|||
fields = ['label', 'coordinates']
|
||||
|
||||
Note that this example doesn't handle validation. Partly for that reason, in a
|
||||
real project, the coordinate nesting might be better handled with a nested serialiser using two
|
||||
`IntegerField` instances, each with `source='*'`.
|
||||
real project, the coordinate nesting might be better handled with a nested serialiser
|
||||
using `source='*'`, with two `IntegerField` instances, each with their own `source`
|
||||
pointing to the relevant field.
|
||||
|
||||
The key points from the example, though, are:
|
||||
|
||||
|
@ -717,6 +718,67 @@ suitable for updating our target object. With `source='*'`, the return from
|
|||
('y_coordinate', 4),
|
||||
('x_coordinate', 3)])
|
||||
|
||||
For completeness lets do the same thing again but with the nested serialiser
|
||||
approach suggested above:
|
||||
|
||||
class NestedCoordinateSerializer(serializers.Serializer):
|
||||
x = serializers.IntegerField(source='x_coordinate')
|
||||
y = serializers.IntegerField(source='y_coordinate')
|
||||
|
||||
|
||||
class DataPointSerializer(serializers.ModelSerializer):
|
||||
coordinates = NestedCoordinateSerializer(source='*')
|
||||
|
||||
class Meta:
|
||||
model = DataPoint
|
||||
fields = ['label', 'coordinates']
|
||||
|
||||
Here the mapping between the target and source attribute pairs (`x` and
|
||||
`x_coordinate`, `y` and `y_coordinate`) is handled in the `IntegerField`
|
||||
declarations. It's our `NestedCoordinateSerializer` that takes `source='*'`.
|
||||
|
||||
Our new `DataPointSerializer` exhibits the same behaviour as the custom field
|
||||
approach.
|
||||
|
||||
Serialising:
|
||||
|
||||
>>> out_serializer = DataPointSerializer(instance)
|
||||
>>> out_serializer.data
|
||||
ReturnDict([('label', 'testing'),
|
||||
('coordinates', OrderedDict([('x', 1), ('y', 2)]))])
|
||||
|
||||
Deserialising:
|
||||
|
||||
>>> in_serializer = DataPointSerializer(data=data)
|
||||
>>> in_serializer.is_valid()
|
||||
True
|
||||
>>> in_serializer.validated_data
|
||||
OrderedDict([('label', 'still testing'),
|
||||
('x_coordinate', 3),
|
||||
('y_coordinate', 4)])
|
||||
|
||||
But we also get the built-in validation for free:
|
||||
|
||||
>>> invalid_data = {
|
||||
... "label": "still testing",
|
||||
... "coordinates": {
|
||||
... "x": 'a',
|
||||
... "y": 'b',
|
||||
... }
|
||||
... }
|
||||
>>> invalid_serializer = DataPointSerializer(data=invalid_data)
|
||||
>>> invalid_serializer.is_valid()
|
||||
False
|
||||
>>> invalid_serializer.errors
|
||||
ReturnDict([('coordinates',
|
||||
{'x': ['A valid integer is required.'],
|
||||
'y': ['A valid integer is required.']})])
|
||||
|
||||
For this reason, the nested serialiser approach would be the first to try. You
|
||||
would use the custom field approach when the nested serialiser becomes infeasible
|
||||
or overly complex.
|
||||
|
||||
|
||||
# Third party packages
|
||||
|
||||
The following third party packages are also available.
|
||||
|
|
Loading…
Reference in New Issue
Block a user