Alternative on-disk storage layout where each column is stored together rather than each row. Typical for analytical/BI workloads: queries touching many rows but few columns are dramatically faster. Because every value in a column has the same type, columnar data compresses well — often ~10× (e.g. 40 TB → 4 TB, fitting on NVMe). Downside: can't really update rows — append-only. Citus exposes this in Postgres via `USING columnar`, and columnar and row tables can be mixed in a single query (cold partitions columnar, newest partition row-based).