Sail ran derived TPC-DS at 3TB and 1TB across a distributed cluster, and the results are the latest evidence for our vision in rethinking distributed compute. Hadoop pioneered open-source distributed processing, built around disk. Spark moved intermediate data into memory, though its shuffle stayed disk-first. When we started Sail, we envisioned another shift in distributed compute, toward doing even more of the work in memory, with an engine built entirely in Rust. We built the first version of Sail's distributed architecture in November 2024, with Arrow Flight carrying shuffle data between workers, and redesigned the cluster control plane earlier this year. Arrow Flight shuffle can look surprising at first. A TPC-DS query with dozens of stages and a few hundred partitions can have more than a million logical shuffle streams active in the cluster. Arrow Flight is built on gRPC and HTTP/2, so one TCP connection per pair of workers carries them all. The same approach has since come up in the DataFusion Comet and Riffle communities as well, and we are glad to have started proving this direction two years ago. The full benchmark ran in 1,289 seconds at 3TB and 444 seconds at 1TB, with no shuffle data written to disk. On similar hardware and settings, Sail outperforms the published TPC-DS results for Spark and Spark accelerators at these scales. The post covers the results, the design, and the full setup: lakesail.com/blog/tpcds-benc…