selene-db v1.3.0

selene-db 1.3.0

selene-db 1.3.0 expands the ISO GQL surface, tightens non-ISO grammar
boundaries, and ships the next retrieval/indexing slice from the development
branch.

Highlights

• Adds typed LET variables, typed session value parameters, leading
  OPTIONAL MATCH, OPTIONAL CALL, direct EXISTS graph patterns,
  correlated NEXT, ISO FOR row expansion, simple CASE operand lists,
  counted ANY path search, and counted shortest-path modes.
• Adds edge property indexes, scoped retrieval filters, BM25 candidate-state
  scoring procedures, and deterministic scan-budget cancellation.
• Hardens keyword-boundary parsing and removes several non-ISO grammar
  shortcuts in favor of strict diagnostics.
• Speeds up compact property maps, property diffs, read pipelines, bitmap row
  filters, BM25 candidate handling, and reciprocal rank fusion.
• Refreshes repository-home metadata for the jscott3201/selene-db namespace
  and keeps the release automation pointed at that repository.

selene-db v1.2.0

Published selene-db v1.2.0.

Crates.io packages:

• selene-db-core
• selene-db-persist
• selene-db-graph
• selene-db-algorithms
• selene-db-gql

The release includes the merged vector/TurboQuant, BM25/text, and JSON work through development #802, plus the v1.2.0 release/package namespace fixes.

selene-db v1.1.0

The marathon "finish what 1.0 should've been" release. Full changelog: CHANGELOG.md [1.1.0].

Highlights

Architecture — single native engine (no extensions)

• Vector index externalized to a separate project; planned full-text/BM25 dropped.
• Procedure-pack system removed; selene-algorithms promoted to a mandatory first-class crate with a native Rust API + one frozen BuiltinProcedureRegistry (5 platform built-ins + 19 algo.*). 6 crates total.

ISO/IEC 39075:2024 conformance

• Cross-type numeric completeness (Int128 / Uint128 / Decimal) end-to-end — GV13/GV14/GV17 now honest; field-name-keyed RECORD equality (§4.15); set-operation arm name-equality strictness (§14.2). Parity invariant held (equality ≡ hash/group key ≡ ordering).
• 3VL analyzer parity, full-numeric negate, RETURN UNKNOWN, honest GQLSTATUS error classes.

Durability & reclamation

• External NodeId/EdgeId ↔ internal RowIndex split, three-step MANIFEST recovery, dedicated audit.log, typed retention + MANIFEST-atomic prune, CORE-internal densify compaction.
• Multi-writer: single-committer thread + WriteTxn::seal(), WAL group-commit fsync-before-publish barrier.

Pre-release deep-review hardening (PRs #202–#215)

• P1 correctness fixes, hot-path performance pass across all crates, dead-surface teardown, and ~150 new tests (2503 → 2655).

Verification

Full release CI green on Linux + macOS: clippy, nextest (all-features), doctests, cargo-deny, cargo-audit, THIRDPARTY currency, file-size, no-secrets, 5-min parser fuzz.

Library-only (no server/wire format — embedder owns transport, auth, TLS). Dual MIT OR Apache-2.0.

selene-db v1.0.0

[1.0.0] — 2026-05-16

First stable release. selene-db is now usable as a Rust dependency for
embedding a property graph engine that targets ISO/IEC 39075:2024 (GQL)
conformance. The public API surface across selene-core,
selene-graph, selene-persist, selene-gql, and selene-pack is
considered stable: subsequent 1.x releases will maintain
backwards-compatible additions and reserve breaking changes for major
version bumps.

Highlights

• Strict ISO/IEC 39075:2024 GQL parser, semantic analyzer, planner,
  optimizer, and executor. The GQL Flagger (Clause 24.6) rejects
  non-standard or unclaimed constructs at parse time. No Cypher, no
  SQL, no SPARQL grammar leaks into the engine.
• In-memory property graph with copy-on-write isolation built on
  ArcSwap + parking_lot::RwLock + imbl persistent collections +
  RoaringBitmap label indexes + typed secondary indexes. Single graph
  write-lock with lock-free reads; strict-serializable transaction
  isolation.
• Write-ahead log (SLDB magic) and rkyv-archived snapshots
  (SLSN magic) with two-step recovery; the persistence crate is
  graph-blind (operates on &[Change]) so graph types can evolve
  independently.
• Procedure-pack registry with JSON Schema 2020-12 manifest
  validation, typestate-sealed activation, and a single mutation funnel
  shared between graph writes and pack lifecycle audit (atomic via the
  WAL).
• Graph algorithm library with 15 algorithms across four families
  (structural, pathfinding, centrality, community), exposed through 19
  algo.* procedures via selene-algorithms-pack. Rayon-parallel
  implementations for APSP, betweenness, PageRank, and triangle count.
• Vector index extension with HNSW and IVF providers, SQ8/PQ/OPQ
  quantization, and 9 vector.* procedures via selene-vector-pack.
• Snapshot-protected runtime surfaces: planner, executor,
  procedure-pack, and algorithm outputs are pinned by golden snapshots
  for drift detection (D21 snapshot harness pattern).
• Engineering posture: #![forbid(unsafe_code)] workspace-wide,
  missing_docs = "deny", 700 LOC per-file cap, rustls-only TLS in
  transitive dependencies, no hand-rolled crypto / TLS / async / serde
  primitives.

Crates

• selene-core — foundation types: Value, IStr interner,
  PropertyMap, LabelSet, schema types, Codec, Origin,
  Changeset, feature_register (ISO feature claims).
• selene-graph — in-memory property graph: storage primitives,
  Mutator write funnel, label/typed/composite indexes,
  IndexProvider extension trait, GraphTypeDef runtime binding.
• selene-persist — WAL format (SLDB), snapshot format with
  TLV-tagged sections (SLSN), recovery pipeline. Graph-blind.
• selene-gql — pest GQL grammar, AST, semantic analyzer, planner,
  rule-based optimizer (13 rules), row-at-a-time executor,
  ProcedureRegistry trait, GQL Flagger.
• selene-pack — procedure-pack registry, manifest validator,
  typestate activation state machine, atomic mutation-funnel audit,
  blake3 content hashing, platform built-ins (selene.health,
  selene.create_index, selene.drop_index, selene.pack.history,
  selene.feature_status).
• selene-algorithms — GraphProjection + ProjectionCatalog
  foundation, four algorithm families. Independent of selene-gql.
• selene-algorithms-pack — procedure-pack adapters that expose
  selene-algorithms through GQL CALL.
• selene-vector — opt-in HNSW and IVF vector indexes with search,
  mutation replay, snapshots, quantization, IndexProvider
  registration.
• selene-vector-pack — procedure-pack adapters for vector search,
  mutation, bulk mutation, IVF search, and IVF stats through CALL.
• selene-testing — shared fixtures, synthetic graph generators, and
  pure-mirror snapshot-harness DSLs. Dev-only.

ISO/IEC 39075:2024 conformance

selene-db targets minimum conformance plus a curated subset of
optional features:

• Mandatory data types: STRING, BOOLEAN, INT, FLOAT.
• Optional types behind ISO feature gates: date/time, decimal, list,
  record, path, references.
• Both GG01 (open graph) and GG02 (closed graph) are supported;
  per-graph choice.
• Default transaction isolation is serializable (Clause 4.6).
• Implementation-defined hooks claimed: IW010 (external procedures
  via CALL), IV011 (dynamic property value type), ID001 /
  IW002 / ID003 (principals, authorization, and privileges as
  embedder responsibilities), IE002 / IE004 (transaction
  isolation).
• No wire format is in scope (Clause 4.2.3 is explicit). Embedders pick
  their own transport.

The feature_register in selene-core is the authoritative list of
claimed Implication-table features. The Flagger rejects unclaimed
constructs at parse time so embedders cannot accidentally rely on
non-standard syntax.

Snapshot and WAL format

• WAL magic: SLDB. Append-only log of Change records with
  configurable SyncPolicy (synchronous, batched, or unsynced for
  testing).
• Snapshot magic: SLSN. rkyv-archived TLV sections; producer-tagged
  for graph core (GRPH), vectors (VECS), quantization (QUNT,
  CQNT, IPQB), IVF (IVF1), OPQ rotation (ROTN), and PQ
  training (PQTC). Extensions register their own section tags.
• Two-step recovery: load snapshot, replay WAL from snapshot's last
  seqno. Selene preserves byte-parity for prior snapshot section
  versions on load while writing the current version.

Performance posture

Benchmarks are local-only via scripts/run-benches.sh (criterion and
iai-callgrind, sequential, with mimalloc as the global allocator).
BENCHMARKS.md is the committed, dated measurement
record. Selected v1.0.0 headlines (Apple M5, 100k-scale full profile):

Surface	Measurement
graph_node_fetch	~2.10 ns (columnar storage, lock-free read)
graph_typed_index_point	~4.53 ns (flat-curve Cow tri-state)
gql_analyze_corpus/m5c	~5.32 µs (full analyzer pipeline)
algo_betweenness 100k parallel speedup	~2.40× over sequential
vector_ivf_search	~2.88 µs per query

See docs/performance.md for the full surface,
tuning knobs, and methodology.

Documentation

This release introduces a full user-facing documentation set under
docs/:

• Getting started
• Embedding selene-db
• GQL reference
• Architecture
• Extension guide
• Vector search
• Graph algorithms
• Persistence and recovery
• Performance
• Contributing

The README is now focused on evaluation and orientation; depth lives in
the documentation pages above.

Stability guarantees

The following surfaces are stable starting with 1.0.0:

• Public types and traits in selene-core (Value, IStr,
  PropertyMap, LabelSet, Change, Codec).
• Public types and methods in selene-graph (SharedGraph,
  Mutator, IndexProvider, write-transaction commit flow).
• Public types and methods in selene-persist (WAL format, snapshot
  format, recovery API).
• Public types and methods in selene-gql (parse, analyze,
  plan, execute_statement, Session, ProcedureRegistry).
• Public types and methods in selene-pack (manifest schema,
  ExternalProcedurePack, lifecycle events).
• Wire formats for WAL and snapshot sections (read-side compatibility
  preserved across 1.x).
• The 13 optimizer rules and their static effects on plans.
• The 19 algo.* and 9 vector.* procedure surfaces.

Platform support

Platform	Status
Linux (x86_64, aarch64)	Primary deployment target
macOS (Apple Silicon, Intel)	Primary development target
Windows	Out of scope

Known deferrals (post-1.0.0)

The following items are intentionally deferred and tracked for future
1.x releases:

• Louvain parallelization (currently sequential).
• Edge-index planner support (typed/composite indexes for edges
  currently return Linear selectivity).
• Analyzer recursion-depth bound (parser is bounded at 64; analyzer
  binding is not yet contractually bounded).
• Mutation/MATCH planner threading of BindingId (currently uses
  per-statement context).
• OPQ rotation inner-allocation tightening.
• Fresh extension crates beyond selene-vector and selene-algorithms.