These are internally measured workloads from 8Braid’s records. The configurations and known differences accompany each result. September source reports and raw logs were reviewed for this page; the runs were not repeated. Download the evidence manifest for source hashes and the scope of each attachment.
Key-value / September 2026Read-only GET over TCP.
An in-memory read test of the 8DB MVCC engine and Redis, with the same connection count and pipeline depth.
8DB · median2.802M/s
Redis 8.10.1 · median2.198M/s
M1 · 8 connections · batch/pipeline 128 · medians of 3 rounds. Populated key counts and hit rates differ.
Workload, settings and interpretation
8 September 2026. MacBook Pro M1, 4 performance and 4 efficiency cores, 16 GB RAM. 8DB binary 2b107249. Read-only GET, 8-byte keys, 100-byte values, 1M-key address space, 5M operations per arm. Both clients use TCP.
8DB uses SimulatedIo. Redis persistence is disabled and its random preload populated about 632K keys, so the hit/miss populations are not identical. This measures in-memory read paths, not durable storage or the cost of every encryption and policy option. The recorded 8DB value checks found zero bad responses.
Indexed person-profile lookup.
A person ID retrieves a profile. This is a simplified IS1-shaped point lookup, with no city join.
8DB · median218K/s
Neo4j 2026.07.1 · best3.41K/s
M1 · 8 workers · one request in flight per worker. Different clients, transports and returned representations.
Workload, settings and interpretation
8 September 2026. MacBook Pro M1, 16 GB RAM. 8DB binary 2aacb496, Rust client over UDS, 100-byte profile blob. Neo4j uses a Python Bolt client over TCP, indexed Cypher returning three properties, 1M Person records, 2 GB heap and 1 GB page cache.
The figure uses 8DB’s median of 218K, 209K and 223K, versus Neo4j’s best of 991, 3,410 and 3,325. Both paths checked returned values. The machine was loaded during the sweep. This is neither a full LDBC score nor a graph-traversal result; the difference includes client, protocol and representation costs.
A radius filter over 6 million trips.
A handwritten Rust spatial kernel was compared with SQL engine paths over synthetic trip coordinates.
8DB prototype kernel14.7 ms
SedonaDB 0.3.0470 ms
DuckDB 1.5.2 + spatial: 140 ms. Query-only means of 3 runs; lower is faster. Historical report, with raw JSON not located.
Workload, settings and interpretation
7 May 2026. AWS m7i.2xlarge, 8 vCPU Intel Sapphire Rapids, 32 GB RAM. Fixed-center radius 0.45 in planar dataset units. The 8DB path uses a bounding-box prefilter and squared-distance scan; the comparators use SQL. Loading and index construction are outside the timed query.
All paths reported 94 output rows; equal counts alone do not establish identical record IDs. The report gives an 8DB branch but no unambiguous binary revision, and its raw JSON was not recovered in this review. This is a historical kernel result, not a current full-database benchmark.
Keep the tradeoffs in view.
A separate September single-write test records RocksDB 11.8.1 at roughly 4.67–4.74 ms p50 and 8DB at 5.54–5.64 ms on the M1. The faster path depends on the operation and configuration.
Continue learningWhat should a database comparison include?
Throughput is one dimension. The workload also determines which data structures, access rules, evidence and deployment requirements the application needs.