Next-Generation Design of TCP actors

[plajjan]

We have basic TCP actors. I'd like to have more functionality. I've been pondering the design for quite some time but haven't really reached a conclusion on something that meets all my requirements.

Requirements:

• support happy eyeball, which is about connecting to a host name with the quickest possible transport
  • for example, we want to connect to example.com, like so TCPConnection("example.com", 80", ...)
  • DNS resolves example.com, for both A and AAAA
  • we get both an IPv4 and an IPv6 address
  • we start connecting to both IPv4 and IPv6 at the same time
  • whichever connection establishes first is used, the other is discarded
• support reconnection in a simple way
  • packets are dropped, it's every day life on the Internet, yet most socket interfaces are very low level - sockets cannot be reused, forcing user to redo everything
    • for example, it is NOT enough to connect once and then pass around a socket, the application must also have address, port etc, so it can reconnect
  • explicit reconnect, but maybe also some option of automatic reconnect?
• work well with capabilities based security
  • we have capability based security, meaning we delegate access to things from a higher privilege actor to a semi-trusted actor, mostly by passing something in the Auth chain, like WorldAuth -> NetAuth -> TCPAuth -> TCPConnectAuth
    • for example, main actor has WorldAuth, creates TCPConnectAuth and passes to HTTPClient actor, which is then able to initialize a TCP connection but cannot do other things, like shut down the system
  • but we can also potentially create the resource itself in the higher privilege actor and pass the resource to the semi-trusted actor
    • for example, main actor directly created TCPConnection and passes to HTTPCLient
    • but this requires that HTTPClient actor, which now owns the TCPConnection can do reconnect, otherwise the connection quickly becomes useless
      • the alternative, soliciting help from the privileged actor to perform the reconnect, does not seem like it would lead to nice code
• reusable for many TCP connections
  • like our base TCP client we have now are for plain TCP
  • but we should have the same design pattern and preferably implementation be reused for OpenSSL or similar protocol that rides on top of TCP
    • OpenSSL needs happy eyeballs too!
• 0 overhead
  • for example, an SSL actor should directly talk to sockets, not be layered on top of a TCP actor
    • we don't want the SSL actor to sit on top of a TCP actor, which would introduce extra data passing, it would imply two worker thread schedulings etc
  • 0 overhead also means we cannot implement happy eyeballs using 2 levels of actors
    • but is it possible with mixin classes?

Thoughts:

• multiple actors, perhaps using mixins or so, to cater to different use cases
  • higher level, simple to use, TCP actor
    • for example does automatic reconnect with resonable backoff
      • this should fit 90% of use cases
  • one more raw & low level TCP actor

[plajjan]

This is pretty much done through #1414. There might be bugs etc but I think we've pretty much nailed the design. CLosing this.