[RFC] Unify model format customize string or Json

[tongwu-sh]

Summary

Currently we have 2 model formats: customize string and json. It is hard to maintain 2 serialize/deserialize for model. Specially for some components with complex parameters like category encoders... (almost 1/3 code is about serialize/deserialize and hard to extend). We'd better only support 1 model format.

Motivation

Json format is more reasonable to me, it is standard format and easy to handle hierarchy structure. Backward should be a big problem, I would suggest for new features we can use DumpToString() == DumpToJson(), and replace old component to use json as default format and keep customize string as deprecate method.

[tongwu-sh]

@jameslamb , @shiyu1994 , @StrikerRUS , @guolinke, @hzy46 not sure your opinion on this? thanks!

[StrikerRUS]

I'm for JSON too! However, I guess we should compare performance of these two formats to make a thoughtful choice.

Note, XGBoost is also using JSON as their main serialization format.

dmlc/xgboost#3980
dmlc/xgboost#5046
dmlc/xgboost#5772

[StrikerRUS]

Linking original issue: #2604.

[guolinke]

I remember these is a multi-threading speed-up for string format model file loading.
If the loading speed of json format is comparable, I am okay with the change.
And we may need to consider backward compatibility.

[jameslamb]

@guolinke 's comments make sense to me, that it would be harder to take advantage of parallelism when parsing JSON than when parsing the existing text format. But I think it's worth trying and benchmarking!

I'll also note here that @lemire, author of fast_double_parser who was so helpful with #3405, has worked on a fast JSON parser that we could consider vendoring in (similarly to fmt, fast_double_parser, `etc.): https://github.com/simdjson/simdjson

[lemire]

@jameslamb We would be there to help if you choose to adopt simdjson. At this point, it is a mature library, with extensive tests and documentation. We support two APIs: our On Demand front-end (high perf. default) and a conventional DOM approach.

[StrikerRUS]

Our friends at XGBoost are trying to adopt Universal Binary JSON format as a binary serialization format: dmlc/xgboost#7545.