DuckDB vs SQLite: 100x more rows on the same $16.49 server

A six-day benchmark compared DuckDB and SQLite as embedded storage engines for Traceway's self-hosted OpenTelemetry stack, running both builds on the same $16.49/month Hetzner CCX13 instance. DuckDB writes metrics 4x faster, spans 3x faster, and logs 15x faster than SQLite, while dashboard queries on all three signals now hold up to 100x more rows at equal or better latency before hitting a timeout wall.

The DuckDB build also ingested a billion metric points in under an hour into just 10.8 GB of disk, a scale the SQLite build never got within 100x of. Early runs crashed under memory pressure until the author added admission control (a concurrency cap plus 503/Retry-After responses); all published numbers come from the fixed backend, with methodology changes and single-shot-run caveats fully disclosed.

For engineers self-hosting observability on modest hardware, the takeaway is that an embedded columnar database can push a single cheap box's ceiling two orders of magnitude further before a client-server OLAP system like ClickHouse becomes necessary.