Benchmarks¶

Reference machine: Intel Core i7-12800H (14C/20T, 2.8–4.8 GHz) · 32 GB RAM · Arch Linux.

Transform throughput¶

A single pipeline on a single-thread Tokio runtime, varying the number of chained transforms. Native is pure Rust (no scripting) and serves as the overhead baseline.

Transforms	native	rhai	lua	python
0	2.39 M	—	—	—
1	1.42 M	488.7 K	209.5 K	21.8 K
3	837.9 K	196.1 K	70.3 K	4.0 K
5	548.3 K	119.8 K	43.2 K	5.9 K
10	327.3 K	61.7 K	21.5 K	—
20	181.5 K	31.5 K	10.8 K	—

Python benchmarks use 1 K envelopes (vs 10 K for embedded runtimes) due to subprocess overhead per invocation. Entries marked — were not measured.

Pipeline scaling¶

Each pipeline runs 2 native + 1 Rhai transform on a multi-thread Tokio runtime. This measures how aggregate throughput grows as more concurrent pipelines are added.

Pipelines	total (env/s)	per-pipeline (env/s)
1	444.6 K	444.6 K
5	1.24 M	247.5 K
10	1.83 M	183.0 K
25	2.20 M	87.9 K
50	2.06 M	41.2 K
100	2.30 M	23.0 K
500	2.41 M	4.8 K

Per-pipeline throughput drops as pipelines compete for CPU. For I/O-bound pipelines (e.g. Kafka source → Kafka sink), the bottleneck shifts to the external system and these numbers do not apply.