ADAPTIVE JOIN EXECUTION IN COMPILATION-BASED EXECUTION ENGINES OF DATABASES

65 pagesQuery compilation and adaptive query processing aim to improve the runtime and robustness of analytical databases. However, due to the high cost of compilation, standard methods for combining these involve shaping the adaptive optimization to allow reuse of a single program instead of recompiling. We combine recent developments in both of these areas to show that both compile-once and recompilation-based execution can be practical for adaptive join ordering. We first introduce a low-latency query compilation framework that manages the trade off between compile time and execution time at all stages. First, we describe abstractions to allow easily generating intermediate representation code. Next, we detail the intermediate representation, backend and optimizations that enable similar execution performance to LLVM while achieving much faster compilation time. We then integrate online join order learning that abandons any a-priori optimization into this framework. We propose two orthogonal approaches: a compile-once approach that uses indirection to permute the join order and a recompilation approach that generates code for each join order. We experimentally compare each approach against optimized, analytical databases (MonetDB, DuckDB) on the join order benchmark, TPC-H and JCC-H. Overall, we find that we are able to match or incur modest overheads on queries unfavorable to adaptive optimization while dominating in queries where optimizers are susceptible to choosing a disastrous query plan. We further show that our low latency compilation framework is able to improve both proposed methods across database sizes and in particular, is critical for practical recompilation-based execution.2023-09-0