October 2026 Summaries

BodoSQL extends Apache Calcite’s planning framework to turn logical SQL plans into optimized physical plans for large-scale parallel execution through a staged pipeline of preprocessing, rule-based rewrites, cost-based search with Calcite’s Volcano planner, and backend-specific physical optimizations. Its custom metadata framework obtains and propagates table statistics such as row counts and distinct-value estimates from sources including Snowflake and Iceberg, while a tuned cost model uses those estimates to compare physical alternatives, particularly for choices such as join ordering and aggregate pushdown. Custom rules improve predicate pushdown, including deriving single-table filters from complex join conditions and locking filters to Iceberg scans for file pruning, with one example improving TPC-H Q7 performance by more than three times. After physical-plan selection, BodoSQL introduces explicit subplan caching, including broader “covering” caches that combine similar computations, and runtime join filters that use build-side min/max ranges and Bloom filters to reduce probe-side data before joins. Future work focuses on more accurate cardinality and selectivity estimates, expanded cost modeling, and applying cost-based decisions to later physical optimizations such as cache-sharing strategies.