Docs GitHub Cloud waitlist

Research

RFCs, design docs,
benchmarks.

Every architectural decision lives in an RFC. Every benchmark is published, including the ones that don't go our way. We work in the open because the engine you depend on should not be a black box. Twenty-one engine RFCs at the time of writing, all shipped and public on GitHub, plus one Cloud-side draft in active design.

The principles

Decisions are written down. Argued in public. Reversed when wrong.
Benchmarks include the dataset, the scale, the hardware, and the raw runs.
The numbers that don't flatter us are published with the ones that do.

Open source Benchmarks

Storage

4 RFCs

How NamiDB lives on object storage: layout, format, read path, adjacency.

Storage engine on object storage.

LSM tree over S3 / GCS / R2 with manifest CAS as the coordination primitive. Conditional writes shipped by S3 in 2024, the last missing piece.

SST format: columnar, Parquet-derived.

The on-disk layout of a NamiDB segment. Columnar property storage, sidecar bloom filters, footer metadata for row-group pruning.

Read path: ranged reads.

Why we serve cold reads with per-row-group GETs rather than full-body downloads. Closes a 14x gap against real-WAN R2 in measured benches.

CSR adjacency: shared across snapshots.

Compressed-Sparse-Row adjacency lists held cross-snapshot in RAM. The reason warm-path traversals land in low single-digit milliseconds.

Query

13 RFCs

Parsing, planning, optimisation, execution. From the Cypher subset through join reorder, factorisation, shortest path, and worst-case optimal joins.

Cypher subset.

The slice of Cypher 25 that NamiDB parses and executes today. The grammar is the contract. What isn't in the subset fails fast.

Logical plan IR.

The intermediate representation every query lowers to. Optimizer passes and rewrite rules operate here.

Write clauses.

CREATE, MERGE, SET, DELETE: how mutating Cypher clauses lower to WriterSession operations against the LSM tree.

Cost-based optimizer.

Cardinality estimation, join ordering, and cost-driven plan selection. The reason IC02 lands in ~1 millisecond warm.

Predicate pushdown.

Push WHERE clauses down to scan operators. Skip row groups whose statistics rule them out. Decode only what survives.

Hash join.

The default join strategy. Build side, probe side, spill discipline.

Parquet predicate pushdown.

Push predicates through the Parquet reader so row groups never decode. Saves the I/O, and the cycles.

Hash semi-join.

EXISTS / IN as a cardinality-preserving semi-join, not a join and a distinct.

Projection pushdown.

Only the columns the query needs make it past the scan. Columnar storage's natural advantage.

Join reorder.

Dynamic-programming join enumeration. The optimizer picks the cheapest order, not the textual one.

Factorization.

Factorized intermediate results that avoid materialising the full Cartesian product. Behind the IC09 win.

shortestPath and allShortestPaths.

Parser, lowerer, and BFS termination on the existing variable-length Expand. shortestPath emits one row per (source, target); allShortestPaths emits every path of the minimum length. Closes the LDBC SNB IC13 + IC14 parser gap.

Worst-case optimal joins via leapfrog triejoin.

A new MultiwayJoin operator that runs Veldhuizen 2014 leapfrog over sorted partner lists. Triangle and clique queries fold from O(d^{k-1}) binary-join intermediates into a single k-way join with an AGM-tight cost model. Feature-flagged behind NAMIDB_WCOJ=1.

Caches

2 RFCs

Cross-snapshot caches that make warm reads cost the price of a hash lookup.

Node view cache: cross-snapshot.

Lookup cache for NodeView reads that survives snapshot boundaries. The second read of the same node hits, even from a newer snapshot.

Edge SST caches.

SST bodies, decoded edge property streams, parsed EdgeSstReaders. The fastest read is the one that never crosses the storage boundary.

Runtime

2 RFCs

How the server fans out reads and what it speaks on the wire. The two unlocks in v0.5: concurrent reads, and Bolt.

Concurrent reads without the writer mutex.

An owned, lifetime-free snapshot held in an atomic Arc cell. Readers stop queueing behind the writer mutex; a single namidb-server fans reads across every core. Measured ~7x throughput at 8 readers on a 4-core box.

Bolt wire protocol.

Bolt v4.4 / 5.0 / 5.4 on a second TCP listener. Every published Neo4j driver (Python, Java, JavaScript, .NET, Go, Rust) connects with bolt://host:7687, unmodified. PackStream codec + state machine in a dedicated namidb-bolt crate.

Next

1 RFC

Drafts in active design. Each lands before the feature does.

RFC-Cloud-1Draft

Gateway control plane: Phase 3-A.

The multi-tenant gateway architecture that fronts the engine in Cloud: wire protocol on /v1/cypher and /v1/bulk, single-writer + epoch fencing between gateway and namidb-worker, scale-to-zero per namespace, and the UsageEvent contract that drives billing.

Draft in design, lands before the feature does

Full archive on github.com/namidb/namidb. Contributions and issues via the same repository.

Run it on your own bucket.

Early access to the Cloud is open. One launch email when the engine is ready, never spam.

Request early access