A fork of GizmoSQL v1.38.0 (a Flight SQL server embedding DuckDB) and of its embedded DuckDB v1.5.5, with server and engine changes:

• GizmoSQL: https://github.com/bddppqs/intent-gizmosql, tag v1.38.0-intent.8 (the release this entry installs, Linux amd64 and arm64); changes in its CHANGELOG.md and CLICKBENCH-FORK.md.
• DuckDB: https://github.com/bddppqs/intent-duckdb, tag v1.5.5-intent.8; changes in its CLICKBENCH-FORK.md.

The scripts are the upstream gizmosql entry's except: install (downloads the pinned release and checks it by SHA-256), query (detects a failed run from the client's exit status and its own error lines, not from result rows), check (SELECT 1), util.sh (the configuration below; start and stop waits bounded at 60 s by the clock) and load (waits polled every 50 ms; the load settings below). The schema and the load SQL are upstream's.

Configuration: on machines with more than 64 CPUs, util.sh sets one DuckDB thread per two CPUs and allocator settings for a large dedicated host; smaller machines run the defaults.

Load: on machines with more than 64 CPUs and one thread per core (as lscpu reports), the bulk insert runs with threads set to the CPU count (DuckDB's default) and zstd_bulk_write_compression_level = 13; both end with the server restart before the first query. Otherwise neither is set; with SMT the insert was slower with them.

Caches: for its lifetime the server keeps decoded string dictionaries, column-wide dictionaries, code translations and their byte lengths, row-group metadata, string-filter outcomes per dictionary entry (and the segments they let a scan skip) and optimized plans of repeated read-only statements. For a statement with a cached plan, the hash-aggregate state is released on a background thread after its result is sent.

Storage: blocks are zstd level 9 with 64 or more threads, 3 otherwise; zstd_block_compression_level overrides it, and zstd_bulk_write_compression_level does for a bulk insert's blocks.

The entry is tuned: yes: for the configuration and load settings above, engine thresholds and three per-architecture build-time defaults (named in the fork documents), and because the changes were developed and evaluated against the 43 ClickBench queries.

Both repositories keep their upstream licenses (DuckDB: MIT; GizmoSQL: Apache-2.0).