Postgres Just Cracked the Top Fastest Databases for Analytics

Original Article ↗ Hacker News Discussion ↗

What pg_mooncake Actually Is

  • Implemented as a Postgres extension providing a columnstore table access method.
  • Data for columnstore tables is stored as Parquet on S3 and local disk, with Delta or Iceberg metadata.
  • Analytical queries on these tables are executed by an embedded DuckDB engine, while Postgres handles catalog, transactions, and WAL (for metadata only).
  • Some commenters argue this means “it’s really DuckDB,” others insist it’s still “just Postgres” from the user’s perspective (psql, extensions, SQL interface).

Performance, Benchmarks, and Scaling

  • On ClickBench it was initially in the top 10 but later slipped to around #12; maintainers say there are easy optimizations left.
  • They claim to be faster than DuckDB on Parquet by storing detailed Parquet metadata in Postgres and using segment elimination to skip files/row groups.
  • Concern raised about Postgres’ linear CPU scaling vs cloud warehouses; maintainers respond that Mooncake’s design allows offloading big queries to “stateless engines” (Athena, StarRocks, Spark, etc.) in future versions.

Roadmap and HTAP / Replication

  • v0.1 has inefficient small writes (one Parquet file per insert); v0.2 aims to fix this and support small-write and time-series/HTAP workloads.
  • Plan: keep OLTP tables in regular Postgres, use logical replication to maintain columnar copies, and run analytics on those.
  • Logical replication is flagged as powerful but non-trivial to run reliably; some warn it can become an entire product’s worth of complexity.

Comparisons to Other Systems

  • Crunchy Data Warehouse: architecturally similar (Postgres + DuckDB + Iceberg); differences cited are open-source licensing, supporting both Iceberg and Delta, and focus on small writes in v0.2.
  • pg_duckdb / Hydra: mainly for querying existing Parquet/S3 data; Mooncake focuses on writing and managing columnar tables in Postgres.
  • TimescaleDB is mentioned as a nearby point on the design space (time-series/columnar) but with different S3 capabilities and licensing.

Deployment, Licensing, and “All You Need Is Postgres”

  • Some are skeptical of calling this “Postgres” given the dependency on a non-core extension and limitations on using such extensions in hosted environments.
  • Legal/licensing concerns around extensions (AGPL, cloud restrictions) are discussed; Mooncake’s extension is MIT-licensed, with assurances it will remain so.
  • Broader thread notes strong demand for “proper Postgres analytics” to avoid complex ETL into separate systems like ClickHouse or BigQuery.