DuckDB Hash Join Optimization
C++ database work on a cache-aligned Bloom filter and software prefetching inside DuckDB’s hash-join operator, evaluated across standard query workloads.
Problem
Hash joins can waste memory bandwidth probing rows that will never match. An optimization needs to reduce that work without adding more overhead than it saves.
Approach
Implemented a cache-aligned Bloom filter and prefetching, then compared strategies across TPC-H, TPC-DS and IMDB. Benchmark results prompted a change from an adaptive-join approach to the Bloom filter design.
Impact
The recorded study reports 1.56× overall TPC-H speedup at 100 GB and covers 135 queries across three workloads. Results depend on the benchmark setup documented in the research write-up; the best selective-query result is separate from the overall figure.
Key Metrics
Technologies
Links
My Role
Worked on the database implementation, benchmark comparison and analysis documented in the research write-up.