Nightly interleaved A/B measurement of XDN's low-load latency (sequential
closed loop, one outstanding request) and max throughput (open-loop rate
ladder) for bookcatalog on Paxos and bookcatalog-nd on primary-backup with the
fuselog statediff recorder on /dev/shm, run on GitHub-hosted runners. The baseline and current HEAD are measured alternately on the same
runner, so the head/base delta is the trustworthy signal; absolute
values vary with runner hardware and are plotted only for context. The SQLite app
runs in WAL mode (ENABLE_WAL=true), and the capacity ladder pins the
primary-backup throughput config (bounded worker pool, PB_INLINE_EXECUTE=false)
so the measured knee reflects code changes, not config drift. Data:
perf.json on the perf-results branch.