Sometimes Filters/DimFilters (Like, Regex, Bound, etc.) on dimensions of very high cardinality (thousands of values) end up unioning bitsets of significant fraction of individual dimension values, and the resulting bitset is 50% or more full.

So instead of making a union of 100k bitsets when filtering on a 200k-cardinality dimension, we better make a bitset of matching dimension value indexes (like in DimensionSelectorUtils.makeDictionaryEncodedValueMatcherGeneric()), scan all rows in the segment and apply a simple check "matchingDimValuesBitset.get(index)" on each row.