• Sources: primary, discussion
  • Summary: Five months after building on per-customer DuckDB instances, PostHog says it could not predict worker sizes for complex queries, and DuckLake writes took over three minutes at about 2 PB, 1 trillion rows and 15.3 million snapshots because commits reread catalog statistics. Serving clients over the Postgres wire protocol was also harder to load balance and keep compatible than HTTP. It is rebuilding on a Trino variant with its own open table format, HogLake.
  • Why it matters: Teams weighing DuckDB or DuckLake for multi-tenant hosted analytics get a first-party account of where single-node engines and portable catalogs stop scaling.