# Changelog
All notable changes to this project will be documented in this file.
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [Unreleased]
### Added
- `FEATURES.md` — honest feature inventory by status, the canonical answer to "what does this crate do?".
- `ROADMAP.md` — long-term direction (async I/O, ChaCha20-Poly1305, pluggable segment store, fuzzing) and explicit non-goals.
- `flake.nix` — reproducible devShell with `zstd`, `pkg-config`, and the Rust toolchain (`nix develop`).
- Shared `benches/support.rs` module consolidating the benchmark `Item`, config, and open helper previously duplicated across all four criterion targets.
### Changed
- Extracted `src/segment.rs`: the on-disk format (filename contract, CBOR→zstd→cipher encode/decode pipeline, segment scan, tmp cleanup) now lives in its own module. `SegmentBuffer` focuses purely on in-memory orchestration and locking. No public API change.
- Renamed the private `BufferInner::pending` field to `unflushed` for precision: it holds items not yet written to a segment file, distinct from the public `pending_count()` backlog metric.
- Crash recovery no longer sorts segments twice or uses guarded `unwrap()`s; the first/last lookup uses explicit pattern matching.
### Fixed
- `delete_acked(acked_seq)` no longer under-reports `pending_count()` when called while items are still buffered in memory. `head_seq` is now clamped to the in-memory window so the backlog count stays honest even when the ack cannot remove unflushed items (they have no segment file to delete). The limitation that acks only take effect on flushed segments is now documented on the method.
## [0.1.0] - 2026-07-19
### Added
- `SegmentBuffer<T>` — durable bounded queue backed by zstd-compressed CBOR segment files.
Generic over any `T: Serialize + DeserializeOwned + Clone + Send + 'static`.
- `SegmentConfig` with tunable batch size, flush interval, max disk usage, and compression level.
- `SegmentBuffer::open(dir, config)` constructor with filename-based crash recovery.
- `append()`, `flush()`, `read_from()`, `delete_acked()`, `latest_sequence()`,
`pending_count()`, `store_pressure()`, `is_overloaded()` public API.
- `SegmentCipher` trait for pluggable at-rest encryption.
- `AesGcmCipher` (AES-256-GCM with random 12-byte nonce prefix) behind the `encryption` feature.
- `SegmentError` (thiserror-based) with `Io`, `Cbor`, `Cipher`, `Integrity` variants.
- 26 unit tests + 2 doc tests covering: basic CRUD, partial reads, limits, crash recovery,
concurrent writers/readers (10K events, 4 writers + 1 reader), time-based auto-flush,
error paths (corrupted zstd, truncated encrypted, wrong key, no key), encryption roundtrips,
and pressure/overload boundary conditions.
### Fixed
- Concurrency bug in `flush()`: sequence numbers (`start_seq`, `end_seq`) are now computed
atomically inside the mutex lock alongside taking the pending events. Previously, a race
between concurrent `append()` calls could corrupt segment filenames by computing the
sequence range from a stale `next_seq` read in a second lock acquisition.
- Same race in `append()`: the returned sequence number is now captured under the same lock
as the push, not re-read after releasing the lock.
### Security
- Extracted from monitor365 and proven on 597M+ events in production.
[Unreleased]: https://github.com/LarsArtmann/segment-buffer/compare/v0.1.0...HEAD
[0.1.0]: https://github.com/LarsArtmann/segment-buffer/releases/tag/v0.1.0