feat: configurable response JSON decoder

[0x12th]

Configurable response JSON decoder

While working on #945, it became clear that request-side JSON serialization and response-side JSON decoding may not belong to the same API surface.

This discussion focuses only on configurable response JSON decoding.

Motivation

The original issue was primarily about customizing how response JSON is decoded, for example using orjson.loads instead of the standard library decoder.

Today response.json() is effectively tied to the stdlib implementation, which makes it difficult to:

• use alternative JSON libraries globally
• optimize JSON decoding performance
• customize decoding behavior consistently across a client

Proposed scope

Only response-side decoding.

No request-side JSON serialization customization.

Proposed API shape

Client-level decoder:

client = httpx.Client(
    json_deserializer=orjson.loads,
)

Per-request override:

response = client.get(
    "https://example.org",
    json_deserializer=custom_decoder,
)

Explicit disable:

response = client.get(
    "https://example.org",
    json_deserializer=None,
)

Proposed semantics

• request-level decoder overrides client-level decoder
• None explicitly disables the custom decoder
• manually-created Request(...) objects without decoder configuration fall back to the client-level decoder
• redirect requests preserve decoder configuration

Internal design direction

Current direction is:

• keep all config merging in build_request()
• propagate decoder configuration through Request.extensions
• configure the Response object during send handling
• avoid introducing new abstractions or generalized serialization hooks

Example internal extension key:

"httpx2.json_deserializer"

Type contract

Current proposal:

JsonDeserializer = typing.Callable[[bytes], typing.Any]

The intention is to keep the contract strict and compatible with libraries like orjson.

Open questions

1. Does the proposed API shape make sense?
2. Should the decoder be configurable at the client level, request level, or both?
3. Is json_deserializer the right naming?
4. Is Request.extensions an acceptable propagation mechanism for this configuration?
5. Should the decoder contract accept only bytes, or str | bytes?

I’d appreciate feedback before continuing the implementation work in the PR.

See also issue #259 for related discussion.

[KeeganOP]

Thanks for putting this together! Ive been giving this some thought. Putting another option on the table, the best I can come up with is the same pattern we use for transports.

Something along the lines of:

import httpx2
import orjson

class MyCustomSerde(httpx2.Serde):
    @staticmethod
    @override
    def serialize(request_object: Any) -> bytes:
        return orjson.dumps(request_object)
        
    @staticmethod
    @override
    def deserialize(content: bytes) -> Any:
        return orjson.loads(content)


client = httpx2.Client(serde=MyCustomSerde)

Which is very close to just passing a function with extra steps but I think encapsulating the two operations would be quite nice.

httpx2.Serde would be a new base class that would just call stdlib json.loads() and json.dumps() (as current) and be created by default in Client but gives consumers a contract and a way to override the default Serde.

Will have to update each method in _api to accept a serde parameter for the request level, and Client.

If consumers don't care about serialize and only deserialize:

class MyCustomSerde(httpx2.Serde):
    @staticmethod
    @override
    def deserialize(request_object: Any) -> bytes:
        return json.dumps(request_object)
        
    # serialize() will fall through to httpx2.Serde.serialize()

For the questions:

1. Does the proposed API shape make sense?

• I think all shapes has up sides and down sides, we will have to find the one that most are happy with

2. Should the decoder be configurable at the client level, request level, or both?

• I would think both. No reason why it shouldn't be

3. Is json_deserializer the right naming?

• Largely based on the outcome of 1

4. Is Request.extensions an acceptable propagation mechanism for this configuration?

• I'm a bit torn. On one hand currently its typed as Mapping[str, Any] so nothing stops adding a key to it. On the other hand I'm nervous for Request.extensions to become a catch all of random things that should actually be supported in a more bespoke implementation. Unless we do see it as the catch all for these types of features?

5. Should the decoder contract accept only bytes, or str | bytes?

• Personally I lean to bytes since that's the current Response.content and the less encoding + decoding the better.