torsteingrindvik / leaning-tower Goto Github PK

The top of the README should more clearly state what kind of problem this library solves.

Right now, futures are boxed wherever necessary.
Following the scheme of tower, we might be able to move this to use custom structs which impl Future instead.

The perf gain (if any) is not really crucial for my own use so this is more of a learning experience.

anyhow is used right now.
Remove it as a dependency and make a "proper" error enum.

If we have a server and we .await it, it will serve the connection as long as communication is successful.
If we return an error on no matching resources (which is a perfectly normal expectation in normal use), the server will quit.

Therefore, we need to change from T into Option<T> and use None where we previously used an Err.

On places in mux_server,
and in Detagger.

We give many services to an allocator in order for it to be able to give us exclusive access of the first available matching service upon requests.

However, we want another layer: We want to allow services to be distributed across endpoints as well.
I.e. we want to be able to have services behind allocators at various IP addresses, but a client shall still be able to ask to see it as a single allocator service.

So:

• Set up several allocators
• Wrap them in some service
  • This wrapper has the same request type as the allocators
  • The wrapper duplicates the request to all allocators
  • The end user will receive the first successful response from any allocator
    • Remaining allocator futures are cancelled

Note that while a single allocator responds with a port number (where the new TCP connection resides), this is not enough now.
The port must be extended to have the IP address as well now.