November 2026 Summaries

After rebuilding its warehouse on DuckDB five months earlier to replace a constrained shared ClickHouse setup, PostHog found that DuckDB’s single-node architecture and DuckLake’s compatibility-oriented design struggled with increasingly large, complex cloud workloads. Although DuckDB performed reliably for local analysis, safely sizing instances in advance, handling memory-intensive queries, and efficiently fetching data from S3 became difficult and costly at scale. DuckLake also created catalog and metadata-management bottlenecks, including slow writes, accumulating obsolete statistics, high database connection costs, and limitations on engine-specific optimization; PostHog’s internal deployment ultimately reached roughly 2 PB, 1 trillion rows, and 15.3 million snapshots. The company additionally found that supporting direct PostgreSQL-compatible connections introduced operational and client-compatibility challenges. In response, PostHog has re-architected its warehouse around a customized version of Trino and a new open table format called HogLake, which it is testing at petabyte scale before expanding access to design partners.