The Risk and Reward of Cross-Engine Query Partitioning: A Case for Partition-Aware Benchmarking

Query engines, whether deployed independently or behind a unified lakehouse interface, evaluate each query on a single engine and report the resulting latency. Whole-query routing treats queries as indivisible units of work, an assumption we challenge both empirically and methodologically. We conduct the first systematic upper-bound study of cross-engine query partitioning, exhaustively evaluating all valid two-fragment partitions across DuckDB and DataFusion on 751 analytical queries. We find that partitioning improves 40% of queries, with speedups up to 14.87×and workload-level gains of 25–33% when compared to the fastest engine. However, the partition space is sharply asymmetric: poor cuts can be orders of magnitude slower, revealing a pronounced risk–reward trade-off. These findings demonstrate that whole-query evaluation systematically underestimates both the performance potential and the risk profile of heterogeneous systems. We propose a partition-aware benchmark protocol that reports three dimensions current evaluations miss: the performance ceiling, the cost of approaching it, and the risk of partition mis-selection.