DuckDB

What it is

DuckDB is an analytical SQL engine that runs inside your own process — pip install duckdb, import duckdb, and there is no daemon, no port and no user account anywhere in the picture. It reads Parquet, CSV, JSON and Arrow directly, so a directory of tick files is already a queryable database before you have loaded anything into its own format.

Three things make it a serious answer here rather than a toy:

• ASOF JOIN matches each row to the most recent row in another table at or before its timestamp — every trade to the quote in force when it printed — producing at most one match per left row, with an ASOF LEFT JOIN variant that keeps unmatched trades. The USING (key, time) shorthand supports only >=; anything else, including the strictly-earlier form, needs the explicit ON condition.
• time_bucket(bucket_width, ts[, origin]) truncates timestamps onto a grid with a chosen anchor, which is bar-building without date_trunc arithmetic.
• Parquet as a first-class storage format, read through globs and Hive-partitioned directories, locally or over S3 through the httpfs extension. Nothing is imported, so nothing is locked in: the same files stay readable by pandas, Polars, Spark or the next engine.

The project is alive on any reading. Commits land on main daily — the most recent at the time of writing was 19 September 2026 — and v1.5.5 shipped on 22 July 2026, the fifth patch on the 1.5.0 "Variegata" line released 9 March 2026, which added a VARIANT type, GEOMETRY as a built-in, a rewritten CLI and non-blocking checkpointing.

Ownership changed in 2026 and the licence did not. DuckLabs, the Amsterdam company employing the core contributors, announced on 26 August 2026 that it was joining AWS as a subsidiary, effective early September; the intellectual property stays with the non-profit DuckDB Foundation and the projects stay MIT.

Pricing

Nothing costs anything, and the FAQ is unusually blunt about it — every component is in the free version and there is no enterprise edition. No node cap, no core cap, no memory cap, no telemetry licence key.

What exists commercially is someone else's product. MotherDuck is a separate, venture-backed company selling managed DuckDB — a free Lite plan with 10 GB of storage and 10 hours of monthly compute, a Business plan from $250 per month per organisation plus metered usage, storage at $0.04 per GB-month and compute from $0.60 to $24.00 per hour depending on instance size. That is a different vendor with a different contract, and buying it is not an upgrade path within DuckDB.

Data & coverage

None. DuckDB ships no market data, no symbology and no corporate actions. It is the engine you point at ticks you already have.

Integrations

Clients for Python, R, Java, Node.js, Go, Rust, C and C++, plus a CLI and a WebAssembly build that runs in a browser. postgres, mysql and sqlite extensions attach live databases; iceberg and delta read lakehouse tables. DuckLake, the Foundation's own lakehouse format — MIT, v1.0 since April 2026 — keeps data as plain Parquet in object storage with metadata in a SQL catalog database, which for tick history means snapshots and time travel without inventing a directory convention. The MCP server for local DuckDB files is MotherDuck's; the Foundation publishes none.

Limitations

• One process writes. Other processes can attach only as READ_ONLY, and then nothing writes at all. The Quack remote protocol lifts this and is documented as beta.
• It is not an ingest target. No line protocol, no Kafka sink, no out-of-order merge on commit — rows arrive because your own code ran INSERT or COPY. A continuously writing feed is the case this loses to a server.
• No server means no server features: no users, no roles, no TLS, no replication, no failover, no backup story beyond copying files.
• Memory is still the constraint on big joins. Larger-than-memory operations spill to a temp directory, but a wide join over a year of quotes can still exhaust RAM in a way a distributed engine would not.
• No market data, no symbology, no adjustment logic. Budget for a feed separately.

Alternatives

ArcticDB if the unit you think in is a versioned Pandas frame rather than a SQL table, and you want a symbol catalogue on S3 without writing one. QuestDB or ClickHouse when a feed writes continuously while something else reads — that is the boundary DuckDB does not cross. kdb+ if the requirement is microseconds and the budget is institutional.

See the rest of the tick data storage category.