Spitballing here, but I’d suspect a “hash table” implemented as a vector of tuples of key/value pairs with linear time search for the keys would outperform a canonical hashtable for simple JRPC APIs where the vast majority of your JSON objects have fewer than a dozen elements.

Would be interesting to build a hashtable with an equivalent “small string optimization” approach where you have a fixed sized array that is replaced by a canonical hashtable once it grows past some threshold.