columnstores

Columnstores are database storage architectures that store data by columns rather than by rows. In a columnstore, values from each column are stored contiguously, often in vertical data blocks. This arrangement favors analytic workloads where queries access only a subset of columns, enabling efficient scans and vectorized processing.

Data layout and compression: Columnar storage exposes strong compression opportunities because values within a column often share data types and ranges. Common encodings include run-length, dictionary, and bit-packing. Since only the required columns are read, I/O is reduced and CPU can operate on compressed data with minimal decompression.

Query performance and updates: Columnstores excel at full-table scans, aggregations, joins on large datasets, and selective predicates. They support predicate pushdown and vectorized execution. However, traditional row-oriented operations such as single-row inserts, updates, and deletes can be less efficient; many columnstores use batch loading or append-only writes, with occasional maintenance to merge or vacuum data.

Use cases and ecosystems: Columnstores underpin modern data warehouses and OLAP databases. They are used by systems such as Vertica, Amazon Redshift, Snowflake, and Google BigQuery, and by file formats like Apache Parquet and ORC in big data stacks. They pair well with compression and distributed storage to scale across nodes and improve scan performance.

Limitations and evolution: While strong for analytical workloads, columnstores may incur higher overhead for transactional workloads and random access to individual rows. Advances include hybrid row-column architectures and adaptive encoding, as well as unified engines supporting both storage models.