Zero-copy Arrow data extraction: we should not be copying / converting the arrow data

[NHDaly]

The goal of the Arrow format is to support zero-copy data transfer over the protocol, so the client SDKs should take care not to copy the data once it's been read from the wire.

However, we don't currently understand how to do this with the Go SDK's API - it's not clear how best to retain the data, nor how best to pass it to users.

[bradlo]

with the updated interface, there is now a single copy, due to the fact that the multipart iterator discards the buffer when moving to the next part. i plan to look further into how to improve this, it should be possible to operate on this data with a single copy directly into userland memory.

[NHDaly]

Note that you won't be able to do it zero-copy yet anyway, since we still haven't implemented the changes in rai-server itself to write each of the parts word-aligned! 😿

So i think 1-copy is the best we'll be able to do in any of the SDKs right now. 💪 great work getting there in the go sdk! :) Thanks Brad!

[NHDaly]

Here's the github issue in raicode for this:
https://github.com/RelationalAI/raicloud-control-plane/issues/2290#issuecomment-1074211947

When/if we moved over to a different transport layer, like Arrow Flight, i think this would be done for us.

[billscheidel-rai]

Note: This issue has been migrated to https://relationalai.atlassian.net/browse/RAI-6340.

This link is only accessible to employees of RelationalAI.