// benchmarks
Numbers you can reproduce.
Every figure on this page comes from the harness in the repo. Clone it, point it at your hardware, and check our work.
12.4 GB/s
In-memory scan throughput, single core
4.1M rows/s
Streaming ingest, sustained
6.4 MB
Static binary, no runtime deps
0 ms
Cold-start time (it is a library)
// tpc-h sf10
Per-query latency
Six representative queries from the 22-query suite. Warm cache, single process, no network.
| Query | Stratabase | Cindergraph | Querium |
|---|---|---|---|
| Q1 — pricing summary | 0.41s | 0.88s | 1.62s |
| Q3 — shipping priority | 0.29s | 0.74s | 1.11s |
| Q5 — local supplier volume | 0.55s | 1.32s | 2.40s |
| Q6 — forecast revenue | 0.08s | 0.31s | 0.59s |
| Q9 — product type profit | 1.12s | 2.84s | 4.97s |
| Q18 — large volume customer | 0.74s | 1.96s | 3.18s |
Geometric mean across all 22 TPC-H queries: 2.6× faster than Cindergraph, 4.9× faster than Querium.
TPC-H scale factor 10 (60M lineitem rows) on a 2024 MacBook Pro M3 Max, 36 GB RAM, warm OS cache. Single process, no network. Lower is better.
// capability matrix
Feature comparison
Throughput is only half the story. Here is what each engine can actually do.
| Capability | Stratabase | Cindergraph | Querium |
|---|---|---|---|
| Runs in-process (no server) | ✓ yes | ✓ yes | ✕ no |
| Vectorized SIMD execution | ✓ yes | ~ partial | ✕ no |
| Zero-copy Arrow | ✓ yes | ✕ no | ~ partial |
| Parquet page-index pushdown | ✓ yes | ✕ no | ✕ no |
| ACID transactions | ✓ yes | ✕ no | ✓ yes |
| Rust / Python / C bindings | ✓ yes | ~ partial | ✕ no |
| License | Apache-2.0 | BSL | proprietary |
How Stratabase compares to two common alternatives for embedded analytics.
Hardware
2024 MacBook Pro M3 Max · 14 cores · 36 GB RAM
Dataset
TPC-H SF10 — 60M lineitem rows, stored as Parquet
Method
Median of 9 runs after 3 warm-up runs, OS cache hot
TPC-H scale factor 10 (60M lineitem rows) on a 2024 MacBook Pro M3 Max, 36 GB RAM, warm OS cache. Single process, no network. Lower is better.