Yuchen Wu

Meanwhile, does it make sense to move `parking_lot` out of the `full` feature group? Other features don't change how tokio works the same way `parking_lot` does.

That should be possible. That is how internally we do the graceful restart. But the code is coupled with signal handling for now. Maybe someone can refactor that code to...

Our load balancing traits are public https://docs.rs/pingora-load-balancing/latest/pingora_load_balancing/selection/trait.BackendSelection.html. @beamandala if you like to implement it, go for it. The code doesn't have to be in our libraries for people to use....

> add a connections field to the Backend struct I imagine you will track the connection info in your own struct. like [ketama](https://github.com/cloudflare/pingora/blob/main/pingora-load-balancing/src/selection/consistent.rs#L23-L27). I don't think the `Backend` itself needs...

For now pingora-proxy is a layer 7 proxy. That means it decrypts and parses the request and response. It also needs to handle connection management (keepalive). [Here](https://github.com/cloudflare/pingora/blob/main/pingora/examples/app/proxy.rs#L46) is an example...

If we ever build a pure L4 relay we will definitely take that into consideration.

`tokio::main` is just syntax sugar for ``` fn main() { let mut rt = tokio::runtime::Runtime::new().unwrap(); rt.block_on(async { // your code }) } ``` So you can do that instead if...

What is your test setup? Benchmark numbers can vary a lot depending on your test setup.

> I tested using release mode and found that one CPU has 2.2W RPS. Does this meet expectations? The raw RPS highly depends on your hardware/OS. To compare, try setting...

> Does it meet expectations? These numbers looks fine to me. Please be mindful that the performance numbers on these synthetic benchmarks are not the direct indicators of real world...