12 hours ago
- DuckDB outperforms SQLite on writes by 4x to 15x and on read cliffs by 100x on a $16/month server.
- Read cliffs (maximum rows for dashboard queries) moved from 1M/100k/100k rows (SQLite) to 100M/10M/10M rows (DuckDB) for metrics, spans, and logs respectively.
- DuckDB stored 1 billion metric points in 10.8 GB on disk and ingested them in under an hour.
- Logs, previously kept off SQLite, became DuckDB's strongest result with 15x faster writes and 100x larger read cliffs.
- A critical ingest bug (lack of admission control) was fixed to produce reliable numbers; stock DuckDB config was used except for increased memory and checkpoint thresholds.
- The DuckDB build is recommended for self-hosting observability unless volume is very low or Alpine compatibility is required.