implementation

Runs the store benchmarks at sizes from 10³ to 10⁵ (or more) on every backend, and prints the Markdown that goes into Benchmark Results. Source: sophia-bench-report.rs.

Sources: Benchmark Plan B-03, B-04, B-05, B-07. code: sophia-bench-report.rs

SOPHIA_BENCH_DATE=$(date -I) cargo run --release -p sophia-bench --bin sophia-bench-report -- --sizes 1000,10000,100000

It reports:

  • the machine;
  • canonical size statistics: median, p90, mean, references per definition, and the share of reference digests in the bytes, with the saving that 8-byte local ids would give;
  • per backend and size: ingest time and rate, size on disk, bytes per definition, the median latency of a record lookup by id, a name lookup and a dependants lookup (1000 random probes), the closure of the last definition, and a full read of the change feed.

B-17 and zero-copy (2026-10-20):

  • --query-size N sets the size for the query workloads (B-17); --skip-queries and --only-queries select them.
  • --zero-copy adds the LMDB read breakdown.
  • B-014 (2026-11-18): the whole-relation workloads (count by kind, unreferenced definitions, names by prefix) also run as “native, tuple at a time” (Tuning::TUPLE_AT_A_TIME: no prefix path, no anti-joins), for comparison in the same run.
  • Bare flags go through flag(). Before that fix, arg() expected a value after every flag, so --skip-analytics never worked.