v0.2.0

What's new in 0.2.0

Atomic transactions

kv.transaction(async (tx) => { ... }) now commits as a single crash-atomic native batch: all staged operations go through one applyBatch native call backed by a new Op::Batch WAL record — either the whole transaction is durable or none of it is. This replaces the optimistic-concurrency retry loop, which could not detect real concurrency in single-threaded JS and threw spurious conflict failures.

Transaction staging is also safer:

• Staged ArrayBuffers are copied at staging time, and tx.getJSON returns a parsed snapshot — mutating your object or buffer between staging and commit can no longer desync what gets written.
• A transaction body that throws or rejects discards all staged writes.

In-memory eviction

maxEntries and ttlSweepIntervalMs now apply to in-memory stores too (previously disk-only). A background sweeper thread expires TTL'd keys and evicts over-limit entries, sharing the eviction-selection logic with the disk writer. The sweeper is stopped and joined on close(), and expired keys are removed under an entry-lock re-check so a concurrent rewrite is never clobbered.

New hook: useKVSelector

Subscribe to a derived slice of a JSON value with an optional custom equality function. Selector and equality functions are kept in refs, so inline selectors don't resubscribe on every render, and object-returning selectors can't cause an update loop.

Durability and correctness fixes

• Batch WAL appends are guarded against concurrent compaction with a per-store RwLock: a snapshot can never capture a half-applied batch and then truncate its record away.
• The sweeper counts live entries exactly, so it no longer over-evicts live keys when more than 4096 keys expire in a single sweep; the WAL sweep deadline uses checked_add to avoid an Instant overflow panic.
• Oversized values now fail fast with a RecordTooLarge error instead of writing a broken WAL record.
• Reopening an already-open store with different OpenOptions (or with recreate: true) returns a configuration error instead of silently reusing the first store — for both disk and in-memory stores.
• encodeBatch falls back to a manual UTF-8 encoder where TextEncoder is unavailable (JSC, older Hermes); batch delete notifies listeners only when the key actually existed.

Example app

New demo cards for an atomic-transfer transaction, namespaced counters, and useKVSelector (with a visible render count), plus self-test coverage for transaction commit, namespace scoping and clearAll, observeJSON, and in-memory maxEntries eviction.

Full Changelog: v0.1.1...v0.2.0

v0.1.1

Docs-only patch: npm badges, MIT license file, and a benchmark chart that renders on npmjs.com (CDN-served SVG). No code changes.

v0.1.0

Initial release: ultra-low-latency, persistent key-value storage for React Native and Expo — a Rust core on scc (lock-free concurrent hash map) behind Nitro Modules.

Highlights

• All reads served from RAM; writes stream to a crash-safe write-ahead log on a background thread (group commits, CRC-protected recovery, atomic snapshot compaction)
• Sync + async APIs, batch setMany/getMany crossing the bridge once
• Encryption at rest (ChaCha20-Poly1305, per instance)
• Per-key TTL with a background sweeper, optional maxEntries eviction cap
• Native change events: listeners and hooks react to writes from any handle of the same store
• React hooks (useKVString/Number/Boolean/Buffer/JSON) and zustand / jotai / redux-persist adapters as subpath exports
• Zero-config persistence (iOS Application Support, Android filesDir), multiple independent stores
• Prebuilt binaries for iOS (arm64 + simulator) and Android (arm64-v8a, armeabi-v7a, x86_64, x86) — no Rust toolchain needed

Install

npm install react-native-scc-storage react-native-nitro-modules

Expo:

{
  "expo": { "plugins": ["react-native-scc-storage"]}
}