Haotian worked on optimizing Parquet metadata handling in the pinterest/ray repository, focusing on improving memory efficiency and stability for large, wide-column datasets. By merging and simplifying dataset metadata within each _fetch_metadata task before transmission to the driver, Haotian reduced peak memory usage and mitigated out-of-memory risks during metadata-heavy data discovery workflows. This approach alleviated driver resource pressure and enhanced the scalability of distributed systems processing large-scale data. The work was implemented in Python and leveraged expertise in data engineering, memory management, and performance optimization, demonstrating a deep understanding of distributed data workflows and the challenges of handling complex metadata.