DuckDB — embedded columnar OLAP, no server

DuckDB is an in-process analytical (OLAP) database: it links into your process like SQLite, but stores data column-by-column and vectorizes execution for aggregates, joins, and window functions. There is no server, no port, no daemon — you pip install duckdb (or drop one CLI binary) and query. Its killer move is reading Parquet/CSV/JSON/Arrow in place, without a load step, so a folder of files becomes a table.

The fact that drives every routing decision below: DuckDB is one writer, many readers, single process. It is brilliant for analysis on one machine and wrong for multi-user serving or transactional app writes.

Latest stable is v1.5.3 (released 2026-05-20). Pin the LTS line (v1.4.x; v1.4.4 LTS shipped 2026-01-26) for anything long-lived — LTS gets ~1 year of patches and a stable storage format. Use 1.5.x for greenfield exploration.

Is DuckDB the tool? (decide first)