“They used a Bloom filter, which is very space efficient and fits in a CPU cache line.”
Researchers at the University of Wisconsin-Madison found single-threaded DuckDB beating SQLite by 30 to 50 times on analytical queries. SQLite only does nested loop joins, and every inner loop probes a B-tree. They added two opcodes that check a Bloom filter before the probe. No hash join, no new planner, and it shipped in SQLite 3.38.0.