danielgtaylor
Integer keys in maps are not parsed correctly from JSON
JSON requires that keys be represented as strings, however proto allows map keys to be any integral or string type. So when specifying a message with a JSON representation, integer keys must be represented as strings. These are not picked up by the .from_json() method, which just runs json.loads(value).
Example: The proto message
message MapTest {
map<string,double> string_map = 1;
map<int32,double> int_map = 2;
}
with JSON example:
{
"string_map": {
"hello": 2.3,
"there": 5.6
},
"int_map": {
"1": 7.9,
"2": 8.3
}
}
Would be parsed by better proto as
MapTest(string_map={'hello': 2.3, 'there': 5.6}, int_map={'1': 7.9, '2': 8.3})
To overcome this we just need to add a custom JSON decoder to the from_json method which looks for any keys which are integers.
An example implementation, adapted from this link would be:
class Decoder(json.JSONDecoder):
def decode(self, s):
result = super().decode(s) # result = super(Decoder, self).decode(s) for Python 2.x
return self._decode(result)
def _decode(self, o):
if isinstance(o, str):# or isinstance(o, unicode):
try:
return int(o)
except ValueError:
return o
elif isinstance(o, dict):
try:
return {int(k): self._decode(v) for k, v in o.items()}
except ValueError:
return {k: self._decode(v) for k, v in o.items()}
elif isinstance(o, list):
return [self._decode(v) for v in o]
else:
return o
We'd then run json.loads(value, cls=Decoder)
An alternative approach would be a type check for map keys in the .from_dict() method, which is then a bit more targeted and should resolve the issue.
I imagine that would hurt performance pretty badly. Not too sure where's best to stand on that.
There are already a lot of type checks in .from_dict(), e.g. for timedelta and datetime. Doesn't seem like adding one for int types, and only if the field type is a map, would have that much of an impact?