ABI Framework nightly Rust
Sample data · not a performance claim

Benchmark dashboard

This page renders data/sample_benchmarks.json — synthetic placeholder records that exist to exercise the dashboard and to give a benchmark exporter a schema to target. No number here was measured from a running ABI build.

Latency percentiles

milliseconds, lower is better

Throughput

operations per second, higher is better
Underlying records

Where real numbers would come from

ABI has two local timing surfaces. Both are same-host development guards, and neither is published as a throughput or latency claim.

abi wdbx benchmark In-process

Times insert and search against the in-process vector index and prints p50/p95/p99 per operation. Its own output header says it plainly: local, in-memory, and not a published throughput claim.

run it 
./tools/cargo.sh build -p abi-cli
./target/debug/abi wdbx benchmark 500

tools/bench_regress.sh Regression gate

Compares the frozen WDBX workload against tools/bench_baseline.json and fails on a slowdown beyond the threshold. It aggregates the best p50 across several runs so host noise can only make the gate more conservative, and it refuses cross-host comparison by default.

run it 
./tools/bench_regress.sh
# ABI_BENCH_COUNT / ABI_BENCH_RUNS / ABI_BENCH_THRESHOLD_PCT
# tune the workload; --update-baseline refreezes it

Publishing measurements here

The dashboard fetches its records at page load, so real data only needs to be written to site/data/sample_benchmarks.json in the documented shape — see site/data/README.md. If that ever happens, the provenance note above has to change with it: which machine, which toolchain, which workload. Undated numbers with no host attached are exactly the kind of claim the claims audit exists to prevent.