- Sources: DuckDB blog, 2026-07-31, HN discussion
- Summary: On TPC-H Q6 at SF100 against S3, DuckDB v1.5.5 kept an average of about 6 of 64 cores busy and stayed near 5 Gbit/s, so the machine waited on synchronous reads. Asynchronous reads cut that query from 8.2 to 2.8 seconds, and the CSV variant from 877.6 to 45.3 seconds. They are on by default from v2.0, scheduled for autumn 2026, and available now in v2.0.0-dev preview builds.
- Why it matters: The bottleneck teams attributed to query execution was request concurrency.
send feedback on this story