Best of DuckDB — March 2026

  1. 1
    Article
    Avatar of motherduckMotherDuck·9w

    DuckDB 1.5 is Faster and Easier than Ever

    DuckDB 1.5 brings major performance improvements and new features. The new VARIANT type enables automatic JSON shredding for 10-100x faster semi-structured data queries. Min/max queries are 6-18x faster via statistics-based optimization, complex joins now leverage hash joins more broadly, and Top-N-by-group queries can be up to 70x faster. Common Subplan Elimination speeds up complex CTEs. New capabilities include reading entire folders of DuckDB files, writing to Azure Blob and ADLSv2 storage, non-blocking checkpointing with parallel WAL files (17% TPC-H throughput boost), and DuckLake 0.4 with macros and sorted tables.

  2. 2
    Article
    Avatar of duckdbDuckDB·9w

    Announcing DuckDB 1.5.0

    DuckDB 1.5.0 ("Variegata") is now available with a major CLI overhaul featuring color schemes, dynamic prompts, a pager, and last-result access via `_`. Key new features include native VARIANT type support (binary semi-structured data inspired by Snowflake/Parquet), a `read_duckdb` table function with glob support, Azure Blob/ADLSv2 write support, and an ODBC scanner extension. The GEOMETRY type moves into DuckDB core, enabling cross-extension geospatial interoperability, WKB storage, shredding compression (~3x size reduction), geometry statistics for query optimization, and CRS type-system support. An experimental PEG parser ships for better error messages and tab-completion suggestions. Lakehouse updates cover DuckLake spec v0.4, Delta Lake Unity Catalog write improvements, and Iceberg table properties. The httpfs backend switches from httplib to curl. Non-blocking checkpointing improves concurrent RW throughput by 17%. DuckDB 2.0 is planned for summer 2026. The v1.4 LTS line continues until September 2026.

  3. 3
    Article
    Avatar of duckdbDuckDB·9w

    Big Data on the Cheapest MacBook

    DuckDB's team benchmarked the entry-level MacBook Neo (Apple A18 Pro, 8 GB RAM, 512 GB SSD, $700) against AWS cloud instances using ClickBench and TPC-DS workloads. In cold runs, the MacBook outperformed cloud instances due to its local NVMe SSD vs. network-attached storage. In hot runs, the large c8g.metal-48xl cloud instance dominated, but the MacBook held its own against a mid-sized c6a.4xlarge. TPC-DS at SF100 completed in 15.5 minutes; at SF300, DuckDB spilled up to 80 GB to disk and finished all queries in 79 minutes. The verdict: the MacBook Neo is not ideal for daily heavy data workloads due to slower disk I/O and limited RAM, but it handles occasional local analytics well, especially when used primarily as a cloud client.

See all DuckDB archives