JSON: support custom encode/decode without intermediate objects

[kevmoo]

The current support for custom encoding (toEncodable param and supporting toJson on classes) and decoding (reviver) deal with JSON literal types (String, num, Map, List, etc) requiring intermediate representations be created for all complex objects.

Propose to add three members (see below) to dart:convert and then update the JsonCodec, JsonEncoder, JsonDecoder classes to support then as alternatives to toEncodable and reviver.

Note: I don't expect devs to hand-write implementations of these types. Instead, code generators – such as json_serializable – could be updated to target the new API.

typedef WriteJson = bool Function(Object source, JsonWriter writer);

// Inspired by SDK code
// https://github.com/dart-lang/sdk/blob/e7bd3edc9f/sdk/lib/convert/json.dart#L513
abstract class JsonWriter {
  void writeObject(Object object);

  void startMap();
  void writeKeyValue(String key, Object value);
  void endMap();

  void startList();
  void writeListValue(Object value);
  void endList();
}

// Note: already exists in the SDK sources
// https://github.com/dart-lang/sdk/blob/e7bd3edc/runtime/lib/convert_patch.dart#L79-L101
abstract class JsonReader<T> {
  void handleString(String value);
  void handleNumber(num value);
  void handleBool(bool value);
  void handleNull();
  void objectStart();
  void propertyName();
  void propertyValue();
  void objectEnd();
  void arrayStart();
  void arrayElement();
  void arrayEnd();
  T get result;
}

Related issues:

• https://github.com/dart-lang/sdk/issues/29750
• https://github.com/dart-lang/sdk/issues/32476

See https://github.com/kevmoo/json_alt

[kevmoo]

CC @rakudrama @lrhn – I've been able to get much faster encode/decode of custom types vs the VM – imagine if someone who knew what they were doing tried!

[zoechi]

cc @davidmorgan would that work well with built_value as well?

[davidmorgan]

Yeah, this could be a significant improvement for the VM.

IIUC for web it'd have to be a parser implemented in js instead of the native one, curious if the slowdown makes it worth it overall.

built_value by default uses lists instead of maps in the serialization format, on the assumption that it's cheaper (CPU/RAM) to allocate lists than maps. It's clearly better to allocate neither, if possible...

[mraleph]

And if we combine this with dart:mirrors and we get a nice generic JSON deserialization support that does not require build step:

https://mrale.ph/blog/2017/01/08/the-fear-of-dart-mirrors.html

[kevmoo]

And if we combine this with dart:mirrors and we get a nice generic JSON deserialization support that does not require build step:

Sure. It's great for both scenarios. All good.

[kevmoo]

Thoughts @lrhn ?

[lrhn]

Definitely possible. Likely something we can add in a package. I don't expect to change the platform JSON parser now (too many things depending on it). Since the platform JSON parser is written in Dart, there is no reason to think this can't be as efficient.

We might get some code sharing by adding it to the SDK libraries.

[kevmoo]

Since the platform JSON parser is written in Dart, there is no reason to think this can't be as efficient.

There are a number of patches that separate the JS and VM implementations. There is also a one-off implementation of parseDouble. It'd be nice if we could build on top of this...and let folks avoid adding a package import.

[kevmoo]

CC @leafpetersen who has been looking at improved JSON (de/en)code

[lrhn]

I'm trying to abstract our current JSON parser into something like this, where users can supply event callbacks.

The first realization is that it will not work using the native JSON parsing of the browser. There is no way to make those parsers provide similar events. Even using JSON.parse with a reviver gives events that are not in a useful order, and it still builds all the maps and lists anyway.

Using a copy our VM implementation (preferably after fixing a few bugs related to integer overflow), it is fairly easy to do the event callbacks, and I have written listeners that build the current JSON structure, or that build a tightly typed JSON structure (where a [1, 2] would be parsed into a List<int>). You can also generate events from a Dart object structure, similarly to jsonEncode, and then pass the events, with the same event API, to a listener that does something with it, like creating a string.

All in all, I think it's fairly pretty, but it does copy the majority of the VM JSON implementation in order to get there. Since we can't get the same functionality natively in the browser, I see no better alternative. Perhaps a solution would be to use the JS JSON parser first, and then traverse the JS structure to generate the events (which still avoids converting the JS structure to Dart maps first).

[rakudrama]

@lrhn Calling JSON.parse and walking the raw JavaScript to generate the events might have better performance than parsing from scratch, and it is always better to have a smaller solution - we don't want all web apps using JSON to grow at all. The callbacks could have an API that matches walking the JavaScript JSON object, e.g. providing additional information, like transferring ownership of the extracted Object.keys list.

I think for the web experience there needs to be a asynchronous version of the API. Parsing 20MB of JSON takes longer that 16ms, so an async version would be necessary to avoid janky apps.

[kevmoo]

FYI: Our friends at Microsoft have a nice layered model here https://github.com/dotnet/corefx/blob/master/src/System.Text.Json/roadmap/README.md

[jonasfj]

CC @sigurdm, @szakarias, this might be relevant when handling jspb in the protobuf package.

[sigurdm]

Yes - exciting - I think an event-based json-parser would enable us to improve protobuf json-decoding (especially on the web).

[lrhn]

Try package:jsontool for something like this. It has a pull-based JSON reader and a generalized JSON sink for writing.

[gnprice]

package:jsontool looks neat! I read the planets example and skimmed the JsonByteReader implementation, and the API makes a lot of sense.

For me the key ingredient I'd need before I could switch to it would be something to generate the readJson methods. (Extra helpful would be to do so from package:json_annotation annotations, or something very similar, to ease the migration.) For a whole fleet of classes with a bunch of fields each, I don't really want to write those methods by hand… and I definitely don't want to review PRs that write them by hand, or that manually keep them updated when modifying the types. So I'd love to see something that does that. Perhaps a macro?