2026-08-22 release · embed · reliability
Embed train v0.2.5 — crash-loop, soak, pinned LSM
f4kvs-ffi v0.2.5 pins f4kvs-lsm v0.3.6. Crash-loop 50× SIGKILL with zero loss, two-hour soak with 10/10 restart anchors. Cluster stays lab.
PugDB is the public name of F4KVS: a Rust LSM you embed, not a cloud you rent. The production path is still FFI + LSM. The server monorepo is the reliability harness. The cluster is lab.
Pins (2026-08-22)
| Component | Tag | Role |
|---|---|---|
| Engine | f4kvs-lsm v0.3.6 | Canonical LSM (hot-path merged: prefix scan, SST pin share, compacted SST reclaim) |
| Embed API | f4kvs-ffi v0.2.5 | C ABI + Go wrappers, cursor pages, prefix scan, pin on lsm v0.3.6 |
| Server | f4kvs-v2 v0.2.0 | Staging, soak, hostile suites — not a generic prod cluster GO |
Owner GO for alpha embed still stands (2026-07-23): memoirs / RAG preprod. NO-GO generic production cluster.
Reliability, not charts
Re-run on 17–18 August after the LSM hot-path landed:
- Crash-loop — 50 rounds × 1,000 ops, killed with SIGKILL, 0 loss (
soak-20260817T200639Z). - Soak 1h + 1h — 1,661,444 cache ops at 462/s, 10/10 anchors after restart, 0 health failures (
soak-20260818T023658Z). - Meso 100k — 100,000 × 4 KB RAG-shaped chunks, 100,050 / 100,050 keys after reopen (July native engine run; integrity gate unchanged).
- Server 20k microbench @
f52790c: put 9,070 / get 29,412 / mixed 37,594 ops/s, 0 errors.
We still publish a small YCSB A/B/D 10k table on /bench, and a larger 432-cell matrix by median (some QL cells lose). We do not headline a 25× PostgreSQL average, and we do not say “faster than RocksDB”.
What this is for
- Embedded storage under self-hosted RAG and memoirs-style apps (C / Go).
- Teams that want reproducible gates (crash-loop, soak, post-restart row count) more than a leaderboard screenshot.
- Not a claim that PugDB powers every Noematic product. Media Cataloger still defaults to Badger for dups/index; PugDB is opt-in there.
Explore the numbers on /bench. Previous note: alpha embed / meso 100k.
— The PugDB team