databend
databend copied to clipboard
Feature: Try Support BloomFilter Collision
Summary We can use runtime filter's bloom index to do collision with parquet block's bloom filter index to do prune in storage level. When do parquet reading, we can improve filter chances.
cc @dantengsky
https://openproceedings.org/2023/conf/edbt/paper-190.pdf for https://github.com/datafuselabs/databend/pull/14970, we find out that in some cases, the false positive is very high, so we can't prune blocks as expected. We introduce BloomRF to solve this which is newer than surf paper.