Repository navigation
v0.1.1
v0.1.1 — Durability, Benchmarks, and Polish
Fixed
- WAL durability — writer now fsyncs after every append by default (
SyncEvery). Configurable viaSyncModefor throughput-sensitive workloads (SyncOnRotate). - Snapshot atomicity — snapshot writes now use write-to-temp + rename to prevent corruption on crash
- WAL replay dedup — duplicate WAL entries (same actor ID + Lamport clock) are skipped during recovery
- Federation merge — fast path for single-shard search results avoids map allocation
Added
DropCollection— remove a named collection and all its data from disk, wired through RPC transportRPCPuller/DBApplier— exported replication adapters for setting up followers without copy-pasting test codeSyncMode/ManagerConfig— configurable WAL sync policy for balancing durability and throughput- Benchmark suite — Put, Search (1K/10K/100K), parallel search, filtered search, History, Open/Close
Benchmarks
384-dim vectors, 2-bit quantization, Intel Core Ultra 9 285:
| Operation | Time |
|---|---|
| Put (vector + WAL + fsync) | 933us |
| Search top-10 (1K vectors) | 82us |
| Search top-10 (10K vectors) | 546us |
| Search top-10 (100K vectors) | 5.5ms |
| Search parallel (10K, 20 cores) | 49us/op |
| Open + Close (1K vectors) | 31ms |