BemiDB is a Postgres read replica optimized for analytics. It is a single Go binary that embeds the DuckDB query engine, stores data in the Apache Iceberg table format (compressed Parquet), and speaks the PostgreSQL wire protocol so existing Postgres clients and tools can query it.

This benchmark pins v0.51.1, the last release of the self-contained, single-binary, local-storage line. The later v1.x releases pivoted to a multi-container product that mandates S3 object storage and a separate catalog database, which does not fit ClickBench's single-VM, reproducible model.

BemiDB has no bulk-import command of its own — its only ingestion path is syncing tables from a source PostgreSQL database. Loading the ClickBench dataset is therefore a two-step pipeline (see load):

1. Bulk-load hits.tsv into a staging PostgreSQL database (COPY ... FREEZE).
2. Run bemidb sync, which reads the table over a serializable read-only snapshot and rewrites it as Iceberg/Parquet in the local ./iceberg directory.

Both steps are included in the reported load time, because together they are what it takes to get the dataset queryable in BemiDB. After the sync, the staging PostgreSQL is stopped so it does not hold RAM during the query phase — only BemiDB is resident, serving queries from its own Iceberg storage.

The reported data size is the size of the ./iceberg directory (Parquet data files plus Iceberg metadata). The staging PostgreSQL is torn down and not counted.

• No Docker is required: the BemiDB binary is downloaded from GitHub Releases (amd64/arm64 auto-detected) and PostgreSQL is installed from the PGDG apt repository.
• Each query is run through the shared driver's cold/warm cycle. BemiDB is a daemon that reads Parquet from the local disk, and the driver stops it and restarts it before each query's first try, so the cold run is a true cold run and no no-cold tag is set.
• Queries and the table schema are the standard PostgreSQL-compatible ClickBench files, unchanged. Any query the engine cannot execute is recorded as null.