Dremel

Two ideas carry the whole system. The first is a columnar storage format that handles nested records: each field is stored in its own column along with repetition and definition levels, so the original structure can be reconstructed without reading fields you didn't ask for. That format is a direct ancestor of Parquet and ORC. The second is a multi-level serving tree that distributes the query to thousands of leaf nodes and aggregates upward, giving answers in seconds instead of MapReduce minutes. The follow-up paper "Dremel: A Decade Later" (2020) describes the move to disaggregated storage and compute.