Columnar databases store each column separately on disk, so analytical queries that access only a few columns out of many skip the rest entirely — reducing I/O by up to 95% on wide tables. This also enables better compression since similar data types are stored contiguously. The traditional trade-off between write-optimized row storage and read-optimized columnar storage historically forced teams to run two separate systems with ETL pipelines between them. CrateDB's hybrid architecture writes incoming data simultaneously to both row-based and columnar stores on the same table, letting the query planner automatically choose the optimal access path: columnar for aggregations, row-oriented for point lookups. This eliminates the need for pre-aggregation, separate data warehouses, or batch ETL steps while supporting sub-second latency on billions of rows.
