Why is header propagation in Rhai faster than via the config?

[lleadbet]

Describe the solution you'd like

I'm working on load testing the various extensibility options available, and noticing an odd discrepancy when comparing a relatively simple config vs. Rhai: Rhai is actually better performance-wise when setting a static header to subgraphs.

Config for reference:

supergraph:
  listen: 0.0.0.0:4040
  path: /
  introspection: true
headers:
  all:
    request:
      - insert:
          name: "source"
          value: "config"

And the Rhai script:

fn subgraph_service(service,subgraph){
    service.map_request(|request|{
        request.subgraph.headers["source"] = "rhai";
    });
}

Data using 1 min load test with 150 rps; order is [min, mean, 50, 90, 95, 99, max]

No customizations (e.g. a baseline):
1.187ms, 4.296ms, 3.615ms, 6.616ms, 7.924ms, 14.763ms, 124.662ms

Config:
1.291ms, 6.471ms, 4.036ms, 10.814ms, 12.345ms, 26.019ms, 513.332ms

Rhai:
1.191ms, 5.026ms, 3.805ms, 8.685ms, 10.404ms, 16.506ms, 223.113ms

As you can see - Rhai is about 2ms faster for p95s than the config.

Describe alternatives you've considered

Config is preferably more performant than Rhai, or at least understood why this is the case.

Additional context

Working to get profiler in place for this, however wanted to file this ticket for tracking purposes.

[Geal]

I can confirm, rhai is faster than the headers plugin. I don't know why yet though