Optimization: Where to parse Parquet to Arrow

[lanlou1554]

Currently we parse the Parquet on the client side, so every time even for cache hit, we have to re-parse the Parquet files again.
How about parsing it on the server side, which enables to cache Arrow result and save one disk I/O (no need to store the file into disk in the client)
But it is a trade-off: more cache space is needed on the server side, and more data needs to be transmitted via the network.