Is your feature request related to a problem or challenge?
Right now ArrowReader only prunes row groups using min/max stats. But for lookups like trace_id = '...' (where values are basically random 128-bit ids), min/max overlaps every row group and nothing gets pruned. The files do have bloom filters on those columns, the reader just never looks at them.
Writing bloom filters already works, you can turn them on per column via WriterProperties. It's purely the read side that's missing.
Describe the solution you'd like
When a predicate has an eq or IN check on a column with a bloom_filter_offset, load the filter and skip any row group that comes back negative. Might make sense as an opt-in on the scan builder, since it's an extra read per row group checked.
#2398 already did this, but it got closed by the stale bot before anyone reviewed it. @dannycjones suggested to file a feature request, so here it is.
A few other places already do this if it helps: Iceberg Java has ParquetBloomRowGroupFilter, DataFusion does it by default (bloom_filter_on_read), and both Trino and DuckDB check bloom filters on read too.
Willingness to contribute
None (edit: I could definitely test it)
Is your feature request related to a problem or challenge?
Right now
ArrowReaderonly prunes row groups using min/max stats. But for lookups liketrace_id = '...'(where values are basically random 128-bit ids), min/max overlaps every row group and nothing gets pruned. The files do have bloom filters on those columns, the reader just never looks at them.Writing bloom filters already works, you can turn them on per column via
WriterProperties. It's purely the read side that's missing.Describe the solution you'd like
When a predicate has an eq or IN check on a column with a
bloom_filter_offset, load the filter and skip any row group that comes back negative. Might make sense as an opt-in on the scan builder, since it's an extra read per row group checked.#2398 already did this, but it got closed by the stale bot before anyone reviewed it. @dannycjones suggested to file a feature request, so here it is.
A few other places already do this if it helps: Iceberg Java has
ParquetBloomRowGroupFilter, DataFusion does it by default (bloom_filter_on_read), and both Trino and DuckDB check bloom filters on read too.Willingness to contribute
None (edit: I could definitely test it)