Hybrid scan lets us split parquet IO into separate stages so we can overlap IO with GPU compute.
This is especially valuable for remote storage (S3) where IO latency is high and being able to get data into pinned host memory ahead of time significantly reduces the time producers spend waiting.
Hybrid scan lets us split parquet IO into separate stages so we can overlap IO with GPU compute.
This is especially valuable for remote storage (S3) where IO latency is high and being able to get data into pinned host memory ahead of time significantly reduces the time producers spend waiting.