Al Brown

A mechanism-level look at how columnar storage works on disk (per-column blocks, zone maps, and predicate pushdown), with ClickHouse MergeTree and Apache Parquet internals side by side.

OLAP is a category of database systems and workloads built for analytical queries across millions to billions of rows. This is what OLAP is, how MOLAP, ROLAP, and HOLAP differ, and why OLAP databases in 2026 run on columnar engines rather than pre-built cubes.

An OLAP cube is a multidimensional data structure that pre-aggregates measures across hierarchical dimensions. Cubes were the standard analytics implementation from the 1990s through the early 2010s; columnar OLAP engines run the same workloads without the pre-aggregation step.

OLAP operations are the canonical analytical actions performed on multidimensional data: drill-down, roll-up, slice, dice, and pivot. Each maps to a specific SQL pattern.

OLAP databases are optimised for analytical queries over millions to billions of rows. Here is how they work, the five categories on the market in 2026, and how to choose between them.

Postgres is an OLTP database, not OLAP. Its row-oriented storage and B-tree indexes are built for point lookups and small writes, not the wide aggregations OLAP workloads run. Here is how each Postgres-for-analytics extension actually performs, and when CDC to a columnar engine is the right call.