Optimizing Pandas Code for Heterogeneous Backends

Data scientists rely on pandas for building data-processing workflows due to its expressive API and rich ecosystem, but Python’s Global Interpreter Lock and single-threaded execution model make pandas a performance bottleneck at scale. Although modern accelerators like GPUs offer massive parallelism, naively offloading pandas workloads often fails to achieve speedups. These failures are primarily caused by inefficient idiomatic code and greedy scheduling strategies that overlook data-transfer overheads. We propose PandaX, a system for automatic optimization and heterogeneous scheduling of pandas workflows across CPU and GPU backends. PandaX uses a multi-agent LLM-based rewriting framework to transform inefficient pandas code into backend-targeted, high-performance equivalents, guided by profiling feedback, correctness validation, and formal equivalence checking. It then applies a learned cost model and a dynamic programming scheduler to assign each code block to the backend that minimizes end-to-end execution time, explicitly accounting for computation and data-transfer costs. To make optimization practical, PandaX performs rewriting and scheduling using small data samples and extrapolates performance to full datasets via learned execution and transfer models. Experimental results on 25 real-world data-science workflows and 22 TPC-H queries show that PandaX achieves up to 7.6x speedup over pandas on real workloads and 10.5x on TPC-H, significantly outperforming existing GPU-accelerated pandas frameworks and bridging the gap between ease of use and high-performance heterogeneous execution.