Python's json module, converts int dictionary keys to strings.

Question

I have found that when the following is run, python's json module (included since 2.6) converts int dictionary keys to strings.

>>> import json
>>> releases = {1: "foo-v0.1"}
>>> json.dumps(releases)
'{"1": "foo-v0.1"}'

Is there any easy way to preserve the key as an int, without needing to parse the string on dump and load. I believe it would be possible using the hooks provided by the json module, but again this still requires parsing. Is there possibly an argument I have overlooked? cheers, chaz

Sub-question: Thanks for the answers. Seeing as json works as I feared, is there an easy way to convey key type by maybe parsing the output of dumps? Also I should note the code doing the dumping and the code downloading the json object from a server and loading it, are both written by me.

Answer 1

No, there is no such thing as a Number key in JavaScript. All object properties are converted to String.

var a= {1: 'a'};
for (k in a)
    alert(typeof k); // 'string'

This can lead to some curious-seeming behaviours:

a[999999999999999999999]= 'a'; // this even works on Array
alert(a[1000000000000000000000]); // 'a'
alert(a['999999999999999999999']); // fail
alert(a['1e+21']); // 'a'

JavaScript Objects aren't really proper mappings as you'd understand it in languages like Python, and using keys that aren't String results in weirdness. This is why JSON always explicitly writes keys as strings, even where it doesn't look necessary.

Answer 2

As bobince said, JSON has no number keys: http://json.org/

Answer 3

This is one of those subtle differences among various mapping collections that can bite you.

In Python (and apparently in Lua) the keys to a mapping (dictionary or table, respectively) are object references. In Python they must be immutable types, or they must be objects which implement a __hash__ method. (The Lua docs suggest that it automatically use the object's ID as a hash/key even for mutable objects and relies on string interning to ensure that equivalent strings map to he same objects).

In Perl, Javascript, awk and many other languages the keys for hashes, associative arrays or whatever they're called for the given language, are strings (or "scalars" in Perl). In perl $foo{1}, $foo{1.0}, and $foo{"1"} are all references to the same mapping in %foo --- the key is evaluated as a scalar!

JSON started as a Javascript serialization (object notation) technology. Naturally it implements semantics for its mapping notation which are consistent with it's mapping semantics.

If both ends of your serialization are going to be Python then you'd be better off using pickles. If you really need to convert these back from JSON into native Python objects I guess you have a couple of choices. First you could try (try: ... except: ...) to convert any key to a number in the event of a dictionary look-up failure. Alternatively, if you add code to the other end (the serializer or generator of this JSON data) then you could have it perform a JSON serialization on each of the key values --- providing those as a list of keys. (Then your Python code would first iterate over the list of keys, instantiating/deserializing them into native Python objects ... and then use those for access the values out of the mapping).

Answer 4

I've gotten bitten by the same problem. As others have pointed out, in JSON, the mapping keys must be strings. You can do one of two things. You can use a less strict JSON library, like demjson, which allows integer strings. If no other programs (or no other in other languages) are going to read it, then you should be okay. Or you can use a different serialization language. I wouldn't suggest pickle. It's hard to read, and is not designed to be secure. Instead, I'd suggest YAML, which is (nearly) a superset of JSON, and does allow integer keys. (At least PyYAML does.)