Strategies for fetching arrow

[surister]

1. partitioned parallel fetch data, same parallel fetch thread also parses rows to RecordBatch
   Note: concat_batches almost doubles the memory usage peak, as a buffer needs to be created to copy all the existing arrays.
2. partitioned parallel fetch data, return Vec<Rows>, then parallelize columns parsing to a single RecordBatch

[surister]

The actual performance gain seems to be to use streaming, loading all the fetched data into memory is not scalable, if we iterate and stream data to arrow arrays as we get them, we use much more less memory, as arrow is very efficient.

[surister]

right now every thread counts how many rows and pre allocates, resulting in total record batches = threads, can we have the allocation done only in ONE thread, and every other thread that gets data writes to that thread?