It's not quite as high as I was seeing with `jiffy` (3,800 req/sec here vs 4,000+ with jiffy), but I'm not confident that was a totally fair comparison. `jiffy` doesn't integrate as nicely with Phoenix, so I was just calling `:jiffy.encode(...)` in the controller and then doing a `text(...)` response. I need to double-check if `json(...)` is doing more work here.
[0] https://github.com/losvedir/transit-lang-cmp/commit/140d693b...
Jsonrs supports encoding protocols by default, but lets you turn it off for a speed boost if you're okay with more Jiffy-like behavior. Plugging it in as Phoenix's JSON library won't turn off that protocol handling, so you're getting the slower operational mode of Jsonrs (which is fine and probably what you want in most cases for correctness, but will definitely eat a few ops/sec in a benchmark)
Personal plug, but have you tried Jsonrs[0]? I typically get much better performance out of it (especially in lean mode, which seems to be the mode you'd want to use for this benchmark) than Jiffy for large JSON encoding workloads.
[0] https://hexdocs.pm/jsonrs/readme.html