Changelog
This page summarises each release of the engine (server, CLI, MCP server,
Python package and container images share one version). The complete,
itemised history, with measurements and internals, is the engine’s
CHANGELOG.md.
Upgrade procedures are in Upgrades.
3.3.0 — 2026-11-03
Added
- Restore points: point-in-time backups.
--restore-point-retention/NAMIDB_RESTORE_POINT_RETENTION(default0, off; for example24hor7d) keeps every committed version of that window restorable.POST /v0/admin/backup(multi-tenant:/:namespace/v0/admin/backup) takesas_of: an RFC 3339 instant, resolved to the newest version committed at or before it, or a manifest version as a JSON integer. The copy is the namespace row for row as of that point, and the response addssource_created_at. NewGET /v0/admin/restore-points(multi-tenant:/:namespace/v0/admin/restore-points, which does not open the namespace) returns the window’searliestandlatestpoints and pages of points, newest first. New error codes:invalid_as_ofandas_of_in_future(400),restore_point_unavailable(404, with the window’s bounds),namespace_not_found(404),invalid_limit(400). Storage grows by what the namespace writes during the window. The server refuses to start with a retention andNAMIDB_COMPACTION_INTERVAL=0s. Each sweep logsrestore_window_retainedandrestore_window_bytes. The CLI gainsnamidb restore-points --from <uri>andnamidb backup --as-of <instant>. See Restore points.
Changed
- Retention leases now carry a scope. A backup pins only the version it copies instead of every version from it to the current one, so backing up an old restore point does not make each sweep load every manifest in between. Older releases read every lease as a floor, which keeps strictly more, so mixed versions stay safe.
Fixed
- The query timeout and the admin cancel now stop every operator
mid-run (#207). A cross
product does no storage reads, so it only checked the deadline between
operators:
MATCH (a:N), (b:N) RETURN count(*)over 3,000 nodes ran for more than 40 s under--query-timeout 1sand ignoredPOST /v0/admin/queries/{id}/cancel. Every loop that works over rows already in memory now checks the deadline and the cancel flag, once per 1,024 rows: cross products, hash joins and semi-joins, uncorrelatedApply, filter, projection,DISTINCT, grouping and aggregation,ORDER BYand top-k, pattern comprehensions, the multiway join, the graph-algorithm projection, andreduce, list comprehensions and list quantifiers. A guardedORDER BYover more than 16,384 rows sorts in runs and merges them, with an identical, stable result. The statement fails with the existing timeout (HTTP 504, BoltTransactionTimedOut) or cancelled (HTTP 409, BoltTransaction.Terminated) error within milliseconds. The cost of the checks is under 2%.
Documentation
- The engine READMEs match the code: Bolt negotiates 4.4, 5.0, 5.4 and 5.7;
NAMIDB_EDGE_POINT_MAX_ENTRY_BYTES/NAMIDB_EDGE_POINT_MAX_SST_BYTESdefault to0(unlimited);NAMIDB_VECTOR_FILTER_MAX_DISTINCTandNAMIDB_VECTOR_FILTER_MAX_BYTESwere never read and are no longer documented;namidb-mcplists all 14 tools.
3.2.1 — 2026-11-02
Fixed
- A running query’s working memory is now bounded
(#203). Admission under
NAMIDB_MEMORY_MAX_BYTESonly sampled memory when a query started, so one admitted query that materialised a large result (a 600,000-rowUNWINDin a 384 MiB container) could grow until the kernel killed the process and every request in flight with it. Each statement, HTTP and Bolt, reads and writes, now runs under a working-memory budget,--query-memory-max-bytes/NAMIDB_QUERY_MEMORY_MAX_BYTES(defaultauto: 50% ofNAMIDB_MEMORY_MAX_BYTES, off when that is0; an exact byte count also works without a ceiling;0disables). All running queries also share a pool of 75% of the ceiling. Over its own budget a query fails with 413query_memory_limit(BoltNeo.ClientError.Statement.ResourceLimitExceeded, GQLSTATUS54000; not retryable unchanged); stopped because the pool is full or the process reached its ceiling, it fails with 503memory_pressure(BoltNeo.TransientError.General.DatabaseUnavailable,53000; retryable). A write whose HTTP response would not fit is rolled back.NAMIDB_QUERY_ROW_CAPstill applies on its own. New metrics:namidb_query_memory_limit_bytes,namidb_query_memory_pool_bytes{kind},namidb_query_memory_stops_total{reason}. See Memory. Upgrade note: a server that setsNAMIDB_MEMORY_MAX_BYTESgets the budget on upgrade; setNAMIDB_QUERY_MEMORY_MAX_BYTES=0to opt out.
Changed
- The 503 for a query refused at memory admission now carries
code: "memory_pressure",neo4j_codeandgql_status, like every other executor error. Theerrortext is unchanged. - The row cap now also applies to a few lookup and
DISTINCTpaths that returned early.
3.2.0 — 2026-10-31
Added
POST /:namespace/v0/admin/restorefor multi-tenant servers: restore a backup into a namespace that does not exist yet, through the server’s own store. Retrying a failed restore is safe; requests that would open the target get 503 while it runs.
Security
--backup-target-uriis enforced on parsed URIs instead of a string prefix: same scheme and bucket, a path at or below the prefix with no.,..or empty segments, and exactly the prefix’s query parameters plusns. Upgrade note: put store options in the prefix and repeat them in requests.
Fixed
- Opening a namespace whose manifest body exists without a pointer (an interrupted backup or restore) now adopts that body.
- Numeric keys are indexed when the constraint or index is declared before the data.
3.1.0 — 2026-10-29
Changed
- Vectors, bytes and number lists are written to the WAL and memtable in binary: smaller WAL (23 MB → 8.4 MB for 2,000 1024-d embeddings) and faster commits and flushes. WAL format v2; flush before downgrading to 3.0.0.
- A bare
LIMITover an expansion no longer re-reads its seeds. - An
OPTIONAL MATCHthat references no outer variable runs once. - Edge iteration (
iter_edge_type,scan_edge_type(limit=)) reads edge files in chunks instead of whole. - Mid-sized label scans reconcile only the label’s ids.
Fixed
- Deleting nodes no longer disables the label postings route until the next compaction.
3.0.0 — 2026-10-29
Breaking
- New node data-file format that 2.x cannot read; downgrades need a backup taken before the upgrade.
CALL algo.*rejects unknown options;algo.shortest_pathyields a fourth column,path.MATCHnever binds one relationship twice;OPTIONAL MATCHmatches its whole pattern or nothing.- Python:
merge_nodesmerges into existing nodes (put_nodesreplaces); bulk rows take the node id from_id.
Added
- Python read-only clients (
Client(uri, read_only=True),refresh(),close()),timeout_ms/max_rowsoncypher,limit=on scans, lazyiter_label/iter_edge_type,sweep(), automatic compaction and sweep in the embedded client. algo.shortest_pathtakes an optionaltargetand returns the route.--fence-policy steal|yieldandNAMIDB_MAX_CONCURRENT_QUERIES_PER_NAMESPACE.- macOS x86_64 Python wheel.
Changed
- Embeddings are stored in binary; projections that do not read them skip them entirely.
- Bolt read transactions no longer take the writer lock; Python reads run without the writer lock and release the GIL.
- Namespace names may contain
_;memory://accepts?ns=.
2.6.x — 2026-09
- 2.6.0: schema introspection on every surface (
db.labels(),db.relationshipTypes(),db.propertyKeys(),db.schema.visualization()); server-side read admission;file://stores on FUSE bind mounts; Neo4j 5 spellings for vector and full-text DDL;--write-timeoutgets its own 60 s default. - 2.6.3:
POST /v0/admin/compactdrains level 0 and blocks; result columns followRETURNorder. - 2.6.10: deleting a connected node without
DETACHreturns 409. - 2.6.14 – 2.6.19:
CREATE INDEXon numeric properties accelerates equality and ranges. - Many correctness and performance fixes for expansions, variable-length paths and anonymous intermediate nodes.
2.5.x — 2026-08
GET /v0/admin/queriesandPOST /v0/admin/queries/:id/cancel.- The tokens file hot-reloads.
- Boolean flags accept
1/0/true/false/yes/no/on/off; multi-tenant mode no longer requires?ns=in--store. - 2.5.1: unprefixed multi-tenant routes with the
X-NamiDB-Namespaceheader; boundedUNWIND range(…)memory.
2.4.x — 2026-08
- Composite indexes (
CREATE INDEX … ON (n.a, n.b)). - Bidirectional
shortestPathfor a bound pair. - Admin backup for multi-tenant servers; group commit for multi-tenant namespaces.
namidb runaccepts;-separated scripts.
2.3.0 — 2026-08
- Group commit (
--group-commit-window). - Multi-tenant Bolt: sessions route by the driver’s database.
POST /v0/admin/backup(online, point-in-time).- Bolt 5.7 with GQLSTATUS metadata.
2.2.x — 2026-08
Breaking
- The server refuses to start without an auth source (
--no-authto opt out). shortestPathendpoints must be bound by an earlierMATCH.'text' + <map/list/node>is a type error; Bolt timeout and limit errors get their own codes.
Added
reduce(…);stdev,stdevp,percentileCont,percentileDisc.- Machine-readable HTTP errors (
code,neo4j_code,gql_status). - The official image runs as uid 65532 and pre-creates
/var/lib/namidb.
2.1.x — 2026-08
- Incremental vector and full-text indexes (base plus delta segments) with globally exact BM25 across segments.
- Range-readable formats for vector, text and node data; a RAM page cache on by default.
EXPLAIN [RAW] [VERBOSE]over HTTP and Bolt; admission gate for concurrent full scans; property DDL on loaded data applies immediately.
2.0.x — 2026-07
- 2.0.0:
DROP INDEX/DROP VECTOR INDEX; Louvain and betweenness; graph projections on everyalgo.*; phrase and prefix full-text queries; demand-driven Bolt streaming; writer health in/v0/health; byte-based flush and write backpressure; bounded writer-lock waits; consistent online backups; compaction off the writer lock; automatic recovery of fenced or poisoned writers. - 2.0.1 – 2.0.6: indexed
MERGE, exact relationship probes, sustained bulk loads,NAMIDB_CACHE_MAX_BYTESas one process-wide ceiling, vector filters applied beforek, the process memory governor and Bolt message admission.
1.x — 2026-06
- 1.5.0:
filteronsearch.vector,search.bm25,search.hybridanddb.index.vector.queryNodes;IF NOT EXISTSon vector and full-text DDL; dimension checks on write; prebuilt server binaries. - 1.4.0:
search.hybrid(RRF and linear fusion),search.vector,db.index.vector.queryNodes, all three metrics from the index, int8 quantization, FastRP. - 1.3.0: the vector index serves Cypher KNN, with filters and freshness.
- 1.2.0: composite unique constraints,
IF NOT EXISTS,SHOW CONSTRAINTS/SHOW INDEXES. - 1.1.0: Cypher DDL (
CREATE CONSTRAINT,CREATE INDEX),CALL { … }subqueries,EXISTS { … },FOREACH, label disjunction, open-ended and parameterised variable-length paths,search.bm25,CREATE FULLTEXT INDEX. - 1.0.0: first stable release; seven graph algorithms.
Before 1.0
The 0.x series (May–June 2026) built the engine: the storage format, the
Cypher subset, the HTTP server and Bolt, the Python bindings, read-your-own-
writes, compaction, vector storage, TLS, metrics, backups and token roles,
and the Markdown vault loader and MCP server. See the full
CHANGELOG.md.