zstd’s approach is fundamentally very different from a basic interning dictionary. Seriously, there’s been tons of research on this. It’s also interesting that Fury claims to be 0-copy and then does all these ad-hoc field compression which is the literal definition of not 0-copy (& yes, varint encoding is not 0-copy and neither is this custom string compression).
Anyway, unless you actually do a proper comparison against zstd with a trained dictionary against a representative corpus, we’re just going in circles with me saying “you need to evaluate your ad-hoc compression against zstd” and you trying to argue from first principles that the custom compression scheme wins.
The thing here are that all those RPC are stateless, so the context for compression are the one message itself. i.e. compress object like `Point(1,2)` only without priori knowledge. Users may use train static zstd. But Fury can't, Fury doesn't know which data will be serialized.
And meta string encoding are not statistical, it won't beats zstd. It's just for cases zstd is not suitable.