Implement reflection cache for unions, tuples, options and maps

[bklop]

Hi Zaid! I recently identified a performance issue when serializing one of our F# domain models. The performance issue was caused by many repeated reflection calls at runtime in the FableConverter.

The "worst case" seems to be large DUs as keys in a Map, with many items in the Map. Our model looks something like this:

type MyUnion =
    | Case1
    | Case2
    ...
    | Case200

type MyRecord {
    Map1: Map<MyUnion, string>
    Map2: Map<MyUnion, string>
    ...
    Map20: Map<MyUnion, string> }

I've made some performance improvements which basically boil down to:

• Caching reflection calls for unions (including options) and tuples
• Building lookup tables for union case names (mainly for deserialization)
• Pre-computing constructors for unions and tuples
• Caching map serialization instances to avoid runtime MethodInfo.Invoke calls

For our domain model (basically the code structure mentioned above) we saw serialization time decrease from 450ms to 1.5ms, which we're pretty happy with!

[Zaid-Ajaj]

This is really cool @bklop 🔥 the code looks good too!

Will try to merge and publish soon 😄

[bklop]

Thanks! But take your time to review it for correctness 😄. I tested it on a proprietary application + I made sure all tests pass (obviously).

[Zaid-Ajaj]

I made a benchmark project just for this, check it out here

The performance is insane 🔥 🔥 🔥

I made a large union (500 cases) and a large record (200 fields of type Map<LargeUnion, string>) and tested the JSON handling against the current JSON library and your implementation:

$ dotnet run
Running operations 10 time(s)
Old (Serialization): 2335L ms on average
New (Serialization): 425L ms on average
Old (Deserialization): 87517L ms on average
New (Deserialization): 4L ms on average
Old (Full convertion): 81919L ms on average
New (Full convertion): 2L ms on average

[Zaid-Ajaj]

PR merged and has been published into Fable.Remoting.Json v2.17 and all downstream packages that depend on it. Thanks a lot @bklop for the awesome contribution ❤️