We generate SQL statements of tens to hundreds of KB (CTE chains and wide UNION ALLs over Parquet) and prepare many of them per publish cycle. Moving from DuckDB v1.5.6 (069cc9f9b5) to v2.0.0 (d54cbbe3df) roughly doubled that cycle, and the whole difference is in statement parse/prepare; catalog work, capture and chunk fetches are unchanged or faster on 2.0.
Everything below is reproducible with the three scripts in this gist. Environment for the micro numbers: macOS 15.3 arm64 (Apple Silicon, 16 cores), Bun 1.4.3, @duckdb/node-api 1.5.6-r.2 vs 2.0.0-r.1, in-memory database, default settings, runs alternated on the same host within the same minute. The host was loaded (load average 26 to 49 during the runs), so absolute times are inflated; the ratios reproduce across runs.
| statement | bytes | 1.5.6 extractStatements | 2.0.0 extractStatements | 1.5.6 prepare | 2.0.0 prepare | |--