2026-07-24 release · embed · reliability
Alpha embed path is ready for memoirs & RAG
Tagged f4kvs-lsm v0.3.1 and f4kvs-ffi v0.2.2, owner GO for the embed path, and product-scale proof with post-restart integrity at 100k chunks.
PugDB (F4KVS) is built for teams that need a modular, Rust-native key-value core they can embed or self-host — with evidence that real workloads survive restarts, not only micro-benchmarks.
What’s new
| Component | Release | Highlights |
|---|---|---|
| Engine | f4kvs-lsm v0.3.1 (superseded — current pin is v0.3.6) | Durable LSM + WAL hardening for long bulk ingest |
| Embed API | f4kvs-ffi v0.2.2 (superseded — current pin is v0.2.5) | C / Go bindings pinned to lsm v0.3.1 |
| Decision | Alpha embed GO (2026-07-23) | Approved for memoirs / RAG preprod dogfood |
The production embed path is f4kvs-ffi + f4kvs-lsm. The server monorepo continues to prove distributed ops in staging and lab scenarios.
Reliability first
Before talking about charts, we care about gates you can re-run:
- 5,500+ automated tests on the server monorepo (nextest workspace)
- Docker staging smoke and join/leave
- Soak with post-restart anchors and zero health failures on the recorded full tier
- QL smoke gate (parser, CRUD, wire path)
- Multi-node scenario suite and hostile fault injection on the cluster path
- Alpha embed GO with pinned tags and a documented checklist
That is the story we want operators to trust: your keys are still there after restart and after messy cluster faults in lab.
Product-shaped scale (100k RAG chunks)
We exercise a RAG-shaped flow — tens to hundreds of thousands of ~4 KB chunks, prefix catalog scans, point reads, and a mandatory post-restart row count.
On the native engine path (no language binding overhead), bulk one-shot ingest of 100 000 chunks lands around ~60k durable ops/s on a developer workstation, with prefix scan of the full key set in a few milliseconds and integrity OK after reopen.
The embed harness (Go bindings) tracks the same integrity gate and similar bulk throughput — so what you measure in-process stays close to what Go services see through the FFI.
Explore the summary on /bench.
Who it’s for
- Embedded storage for self-hosted RAG and memoirs-style apps (C / Go via FFI)
- Self-hosted KV with modular engines and a serious test culture
- Teams that prefer reproducible proof over anonymous leaderboard screenshots
Get started from the tagged embed train as of this note: f4kvs-ffi v0.2.2 + f4kvs-lsm v0.3.1. Current pins: see embed train v0.2.5.
— The PugDB team