The hot path inside a single edge worker is small enough that you can put the entire request lifecycle on one page. That's the trick. Everything else is about keeping the page short.
The shape of a request
A request arrives over QUIC, gets parsed once, picks up a route, and is handed
to a WebAssembly module that gets exactly one allocation budget and one
deadline. If it returns within budget, we forward upstream. If it doesn't, we
short-circuit with a 503 and a via header.
Nothing on that path touches Go's garbage collector. The QUIC stack is C,
the router is a flat table, and the Wasm runtime is wasmtime configured to
preallocate its slab.
Where we cheated
We do not enforce isolation per-tenant inside a worker pool. We enforce it at the worker boundary. That's a deliberate trade: it makes the hot path ~3x faster, at the cost of making blast-radius math harder.
What we'd warn you about
If you benchmark only the steady state, you will be disappointed when real traffic shows up. The interesting numbers live in the tail.