Fig 1The native storeHover to slow down · Space pausesframe 128 · 4 waiting for meta/lock · 2048 objects, no lock
The commit in flight: the one writer holding meta/lockCommits waiting for the lock, in no strict order, and objects dropping into fan-out bins with no lockWriters' metadata transactions wait for one lock, an flock: at each release the kernel wakes every waiter and the first to run takes it, so the order is not strict and a later writer can go first. Each transaction becomes one CRC-framed frame appended to journal.log; objects are made durable before any frame names them, into 256 fan-out directories, without a lock (sixteen bins drawn). Measured on one busy host (docs/perf.md, 2026-10-16, with the writer lock that ships): with 32 processes committing through the binary at once, 72.7 commits a second and a median commit of 427 ms (p99 656 ms), about 406 ms of it the wait for the lock; the metadata transactions alone, at store level, 120.9 a second; and object writes on 32 threads in one process 13.8× as fast as on one thread, as the disk and ext4's journal commit concurrent fsyncs together.