Local spend benchmark#

hubu-bench simulates multiple agents owned by one user submitting spend requests to a local Hubu server at a fixed rate. It checks both performance and the consistency of authorization, budgets, mock payments, and ledger state.

Run#

./scripts/benchmark-local.sh

The script builds hubu-server and hubu-bench, starts an isolated server on 127.0.0.1:8790, samples CPU and RSS, and writes results beneath target/hubu-bench/.

Common overrides are:

HUBU_BENCH_AGENTS=8 \
HUBU_BENCH_RPS=16 \
HUBU_BENCH_DURATION_SECONDS=20 \
HUBU_BENCH_WORKERS=8 \
./scripts/benchmark-local.sh

For direct use:

cargo run -p hubu-bench --bin hubu-bench -- \
  --url http://127.0.0.1:8787 \
  --agents 4 \
  --rps 8 \
  --duration-seconds 10 \
  --workers 4

Correctness gates#

The benchmark fails when a configured guardrail is violated. It verifies:

planned requests were recorded;
request errors remain below the configured limit;
allowed decisions receive authorization tokens;
successful payments settle holds;
failed payments release holds;
successful payment and ledger counts agree;
consumed budget equals successful payment spend; and
no budget remains frozen at the end of the run.

Interpretation#

The benchmark measures the local HTTP path through policy, authorization, budget reservation, mock payment, settlement, and ledger recording. It is not a capacity claim, service-level objective, production load test, or security test. Results vary by machine and should remain generated artifacts rather than being copied into this runbook.

Use step-load runs and preserve the generated report when comparing a specific commit. Open product or implementation gaps in the issue tracker instead of maintaining a dated backlog in this document.