[morganherlocker]

If most data between the backend/frontend is raw numeric array binary blobs, and you target array types that can be loaded directly into JS TypedArrays (int32_t* -> Int32Array, float* -> Float32Array, etc.), you have a more or less optimal data transfer mechanism with zero dependencies. The JS parsing is faster than the fetching, even with 10s of millions of items, provided you keep the number of arrays fixed and reasonably small.

This is a natural fit for data-oriented programming in most systems languages, including C++. It's also compatible with numpy and any other dynamic language backends that have a performance-oriented array primitive or lib. The data can be served dynamically from memory, or written to files on a static server like nginx.