NamiDB
ProductResearchOpen sourceCompanyRequest early access
Benchmarks

Every number,
with its caveat.

Below are the measurements that closed the Fase 1 storage gate and the Fase 2 query gate. Four LDBC SNB Interactive queries against Kùzu on SF1 + SF10. Six storage benches against LocalStack and the first real-WAN run on Cloudflare R2. Same hardware. Every caveat in the open.

Query engine — LDBC SNB

  • IC02Faster

    Recent messages from your friends

    22× faster than Kùzu on SF1.

    0.00ms
  • IC07Faster

    Recent likers of your messages

    1.4× faster than Kùzu on SF1.

    0.00ms
  • IC08Faster

    Recent replies to your messages

    55× faster than Kùzu on SF1.

    0.00ms
  • IC09Faster

    Friends of friends, with their messages

    3.5× faster than Kùzu on SF1.

    0.00ms

Four of four queries pass the 2× Kùzu gate on both SF1 and SF10. NamiDB is faster than Kùzu on seven of the eight (query, scale) cells. The remaining ten IC queries land in the next milestone. We will publish what we measure — including the ones that go the other way. Raw runs and reproduction scripts live in docs.namidb.com/bench.

Storage engine

Fase 1 gates,
closed.

The storage layer sits on object storage with manifest CAS as the coordination primitive. Below are the numbers that closed the Fase 1 plan — cold-read, warm-read, ingest throughput, recovery, concurrent mix, and the first real-WAN run against Cloudflare R2.

  • READ · coldPass
    383 msGate < 500 ms

    Snapshot lookup_node, cold cache

    LocalStack · 100 M nodes

    Closed with 113 ms headroom. Constant in total dataset size — bounded by one SST decode, not by the namespace.

  • READ · warmPass
    9.4 msGate < 10 ms

    Snapshot lookup_node, warm cache

    LocalStack · 100 M nodes

    Closed after the sorted-by-min-key descriptor index landed. Manifest scan stopped dominating at 10 SSTs.

  • READ · WANPass
    220 msGate < 500 ms

    Cold lookup_node with ranged reads

    Cloudflare R2 · 1 M nodes

    First real-WAN measurement. RFC-003 ranged reads + metadata cache land a 2.14× win over the full-body GET that LocalStack hid.

  • INGESTPass
    281 K /sGate ≥ 10 K /s

    WriterSession end-to-end nodes/s

    LocalStack · 1 M nodes

    28× headroom. Inmemory floor is 480 K /s. Parquet decode adds ~0 % overhead — fully amortised against memtable apply.

  • RECOVERPass
    110 msSnapshot

    WriterSession::open replay

    LocalStack · 10 × 10 K WAL

    Replay throughput 910 K elem/s. Even the adversarial 1 M-record scenario recovers in ~1 s. ~10× cheaper per record than the original write.

  • MIXPass
    6.7 msGate < 10 ms

    Reader p50 under writer pressure

    LocalStack · 15 s window

    p99 9.5 ms. Read path is robust under concurrent writes — cache hit rate ≈ 100 %. Manifest reload dominates the p999 tail.

LocalStack runs on loopback (sub-ms RTT, no TLS). R2 runs over a residential laptop link (~200 ms RTT, HTTPS, TLS via aws-lc-rs). Real S3 same-region from EC2 sits between the two. Total R2 session cost: < $0.02. The full per-bench writeups — including the runs that failed and what the fix was — live in nitrograph/docs/benchmarks/.

Build the next decade with us.

Early-access is open. One launch email when the engine is ready — never spam.

Request early access