gg2

DuckDB async I/O internals

Archived — this story has rotated out of today’s deck. It is kept here in full.

The gist

DuckDB v2.0, due fall 2026, adds async reads for Parquet and CSV.
Async I/O cuts remote query times dramatically, like TPC-H Q6 from 8.23s to 2.84s.

Background

DuckDB, an in-process analytical database, traditionally ran on local SSDs with synchronous I/O. As it expands to cloud data lakes (e.g., S3), synchronous reads block worker threads, wasting bandwidth. To fix this, DuckDB is implementing asynchronous I/O with separate thread pools and read-ahead, targeting Parquet and CSV first.

How it unfolded

  1. May 2026DuckDB introduced the Quack protocol, enabling DuckDB to run as a server, expanding remote use cases.
  2. Jul 31, 2026DuckDB published a blog post detailing async I/O internals, announcing v2.0 preview builds.
  3. Aug 16, 2026The blog post trended on Hacker News; MotherDuck highlighted performance gains, e.g., TPC-H Q6 on S3 from 8.23s to 2.84s.
  4. Fall 2026DuckDB v2.0 is scheduled for release with async I/O enabled by default for Parquet and CSV.

Who’s saying what

Official
DuckDB states async I/O can significantly speed up queries when synchronous I/O doesn't saturate bandwidth, as in EC2/S3 setups.
Analysts
TipRanks notes the async I/O improvements could strengthen MotherDuck's positioning in modern data infrastructure.
Community
Hacker News commenters discuss speculative parallel CSV parsing as a related technique, showing interest in internals.

Still unverified

The specific claim that a CSV query dropped from 878s to 45s is mentioned in MotherDuck's LinkedIn post but not detailed in the blog; exact benchmark conditions are unverified.

Sources

See today’s stories in the app gg2 — free on the App Store