Query-Only Architecture

canardstack is one synchronous Rust process backed by DuckDB. It attaches to a DuckLake catalog that another DuckDB process writes through duckdb-otlp, then serves bounded compatibility query APIs over the visible tables.

The query-only boundary keeps canardstack small and predictable. Ingest, durability, catalog serving, and table maintenance belong to the DuckDB writer process and the catalog deployment chosen around DuckLake. canardstack only needs to attach the catalog, validate the schema version, and run bounded reads.

This split also makes the compatibility APIs intentionally conservative. Prometheus, Loki, and Tempo clients expect interactive queries, but DuckLake is the durable storage layer. canardstack therefore favors explicit time ranges, row limits, request timeouts, memory limits, and query admission over broad unbounded query semantics.

Direct SQL remains outside the HTTP API. Use DuckDB, MotherDuck, or another SQL client when you need ad hoc analysis over the tables.

For a deeper implementation map, use the repository architecture docs.