# bison-db -- Roadmap
> Path from scaffold to a stable 1.0. Hard parts are front-loaded; each phase has hard exit criteria.
>
> **Anti-deferral rule:** no listed hard task moves to a later phase unless this file records the move and the reason.
---
## v0.1.0 -- Scaffold (DONE)
Compiles, CI green, structure correct, no domain logic.
- [x] Manifest, README, CHANGELOG, REPS, dual license, CI, deny, clippy, rustfmt, FUNDING.
- [x] API surface sketched in `docs/API.md`.
---
## v0.2.0 -- document model + single-file store + insert/get/delete (DONE)
Document model (`Value`/`Document`), versioned CRC-checked log-structured
single-file store, and `insert`/`get`/`update`/`delete`/`flush` with replay-based
crash recovery. Optional `serde` for the document model.
Exit criteria:
- [x] Every public item has rustdoc + a runnable example.
- [x] Core invariants property-tested (lossless round-trip; index matches file
after reopen).
---
## v0.3.0 -- secondary indexes + field and range queries (DONE)
Ordered secondary indexes over document fields (any number per store), with
equality (`find`) and range (`range`) queries. Both queries also work without an
index via a full scan, so the index is a pure speedup. Indexes are in-memory and
rebuilt per session (`create_index`), keeping the on-disk format unfrozen.
Exit criteria:
- [x] New surface tested (index/scan parity property-tested); hot paths
benchmarked (indexed point lookup vs full scan).
---
## v0.4.0 -- durability policy + crash recovery + on-disk format freeze (DONE)
bison-db is log-structured, so the data file already provides write-ahead
durability; rather than bolt on a redundant second log, v0.4.0 makes durability
*configurable* (`SyncPolicy::Always` for per-write `fsync`, `Manual` for
flush-controlled), hardens recovery (parent-directory `fsync` on create,
best-effort sync on drop), and **freezes the on-disk format** (version 1, spec in
`docs/FORMAT.md`). Files from 0.2.0 onward stay readable by every later release.
Exit criteria:
- [x] No `todo!`/`unimplemented!`. On-disk format frozen and specified.
---
## v0.5.0 -- concurrency + compaction + API freeze (DONE)
`Db::compact` reclaims the space left by overwrites and deletes via a crash-safe
temp-file-plus-atomic-rename swap (verified on Windows and Linux). Concurrency
follows a single-writer, multi-reader model: `Db` is `Send + Sync` (asserted at
compile time) and shares across threads behind `Arc<RwLock<Db>>` — many readers
or one writer, enforced by `&self`/`&mut self`. A single append-only file has one
tail, so single-writer is inherent and intended, as in an embedded SQL engine.
### Frozen public API (as of v0.5.0)
The surface below is **stable**: additive changes only until 1.0, and no breaking
change before then. The authoritative item-level reference is `docs/API.md`.
- Types: `Db`, `DbOptions`, `SyncPolicy`, `DocId`, `Stats`, `Document`, `Value`,
`Error`, `Result`, and the constant `MAX_RECORD_BYTES`.
- `Db`: `open`, `open_with`, `insert`, `get`, `update`, `delete`, `contains`,
`len`, `is_empty`, `ids`, `flush`, `compact`, `path`, `stats`, `sync_policy`,
`create_index`, `drop_index`, `indexes`, `find`, `range`.
- `DbOptions`: `new`, `sync`, `build_sync_policy`, `open`.
- `Document`: `new`, `with_capacity`, `set`, `get`, `get_mut`, `contains_key`,
`remove`, `len`, `is_empty`, `clear`, `iter`, `keys`, `values`, plus
`FromIterator`/`IntoIterator`.
- `Value`: variants and accessors (`is_null`, `as_*`, `type_name`) plus the
`From` conversions.
Exit criteria:
- [x] Public API frozen (recorded here). `cargo audit` + `cargo deny` clean.
---
## v0.6.0 -- Alpha: hardening (DONE)
API and format are already frozen (v0.4.0 / v0.5.0), so this phase is about
confidence, not features.
- In-tree fuzzing of the parse/recovery paths: property tests proving the decoder
and `Db::open` never panic on arbitrary, corrupted, or truncated input.
- Captured benchmarks with method and environment recorded in `docs/PERFORMANCE.md`.
- A realistic example (`session_store`) exercising indexed lookups, per-write
durability, and compaction together.
---
## v0.7.0 -- Beta: soak testing (DONE)
- Seeded stress/soak test driving thousands of randomized mixed operations
(insert/update/delete/read/compact/reopen) against an in-memory model, with
secondary-index parity checked throughout.
- Error-path coverage: oversized-value rejection leaves the store unchanged;
compaction preserves the sync policy.
- Architectural competitive analysis recorded in `docs/PERFORMANCE.md` (the
controlled peer head-to-head is deferred to the RC, to avoid a heavy benchmark
dependency in the library).
---
## v0.8.0 -- Release candidate: head-to-head (DONE)
- Controlled head-to-head benchmark against `redb` (pure-Rust ACID embedded KV),
in a standalone `benchmarks/` crate detached from the package so the peer
dependency never touches bison-db's tree, audit surface, or CI. Result (100k
records, matched durability): bison-db ~1.85x faster on bulk insert and ~35%
smaller on disk; redb ~1.3x faster on point reads. Recorded with full method in
`docs/PERFORMANCE.md`.
- Known read-side optimization avenue (read cache / memory-mapped reads)
documented for a post-1.0 release; not a blocker, since the API and format are
frozen and the gap is modest.
---
## v0.9.0 -- Final soak: concurrency under load (DONE)
- Sustained-load concurrency soak: four reader threads continuously read and
query a shared `Arc<RwLock<Db>>` while a writer runs 3,000 mutations
(insert/update/delete/compaction). Reads never panic under a changing store,
and the store matches the writer's model exactly once the writer finishes —
exercising compaction and mutation under concurrent reads.
This is the last hardening step before 1.0; the API and on-disk format are frozen
and unchanged.
---
## v1.0.0 -- Stable (DONE)
The stable release. Every roadmap item is delivered: document model, single-file
store, secondary indexes with field and range queries, configurable durability,
crash-safe recovery, space-reclaiming compaction, and a single-writer/multi-reader
concurrency model — fuzz-tested, soak-tested, and benchmarked head-to-head.
The public API and the on-disk format (version 1) are the **stability
commitment**: no breaking change until 2.0, per semantic versioning. Files
written by 0.2.0 onward remain readable.
---
## Out of scope for 1.0
- Networked/server mode - bison-db is embedded only; a server is a separate product.
- Graph and vector models - those are other Bison / AnimusDB family members.
- Distributed storage / replication - out of scope for the embedded core.