Incremental updates: only the rows a change affects are recomputed
Shared structures: filtered and joined views reuse their source's column data
Chunk processing: orders of magnitude fewer method calls, and SIMD processes several values per instruction