Improve `SELECT text` performance

[kou]

It's slower than the PostgreSQL protocol.

The followings may be related but we need to look into it:

• Building arrow::RecordBatch is slow?
  • We need to build arrow::RecordBatches to use arrow::ipc::RecordBatchWriter(). We need to copy PostgreSQL data for it. (Not zero-copy.)
  • Should we add an API that writes Apache Arrow streaming format data without building arrow::RecordBatch to Apache Arrow C++?
• Calling SPI_getbinval() is slow?
  • It calls nocachegetattr() https://github.com/postgres/postgres/blob/3edc6580c0e27fb8f13322efd255a88d20dda6c2/src/backend/access/common/heaptuple.c#L496-L712 and it's not a short function. Can we shortcut some operations?

[davidhcoe]

I was wondering if there is an update on this for the performance comparison. And whether any research had gone in to whether this might work with https://github.com/timescale/timescaledb.

[kou]

I have some more ideas (e.g. internal naive ring buffer implementation may be bottleneck) but I haven't worked on this so much yet.

I haven't tried this with TimescaleDB but it will work because this doesn't care about executed SELECT.