Work, Thread, Work — A New Paradigm for Database Performance

Asynchronous I/O in DuckDB

Based on DuckDB Engineering Blog

July 2026

02

Why Async I/O Matters

Breaking the synchronous bottleneck in modern databases

schedule

Threads wait idle during disk or network operations, wasting valuable CPU cycles

trending_down

Adding more threads doesn't help when they're all waiting on I/O

speed

Each I/O operation adds delay, compounding into significant query slowdown

1

Submit I/O request without blocking — thread continues processing

→

2

Thread moves to other tasks while I/O completes asynchronously

→

3

Resume original work when I/O finishes, minimizing idle time

05

The mechanics behind DuckDB's asynchronous I/O model

01

Task Queue Design Work items queued and scheduled across available threads dynamically
02

Non-blocking Submission I/O requests return immediately, allowing threads to stay productive
03

Completion Callbacks Efficient notification system signals when I/O operations finish
04

Work Stealing Idle threads can take work from busy threads, maximizing utilization

85%

Thread Utilization

60%

Query Speed

72%

CPU Efficiency

55%

I/O Throughput

Performance gains measured in analytical query workloads vs. synchronous baseline

Asynchronous I/O: Making databases faster by keeping threads busy

https://duckdb.org/2026/07/31/asynchronous-io