Reliable AutoNat

[jacobheun]

Design notes

The AutoNAT service was written before libp2p had an event system and hasn't seen much love since. If we're going to rely on AutoNAT to reliably detect NATed nodes, we'll need to shore it up a bit:

• Instead of waiting timeouts, we should re-test when our addresses change. Re-check our NAT status when addresses change go-libp2p-autonat#34
• Address all the issues in libp2p/go-libp2p-autonat.
• Turn on a rate-limited AutoNAT service on all (DHT?) nodes by default. This is especially important for when we start persisting the routing table to disk and bootstrapping off that persisted routing table: we're unlikely to randomly run into an AutoNAT service node.
• We should try to advertise fewer junk addresses, either by confirming our addresses with AutoNAT or some other means.

Testing mechanics

This will likely involve quite a bit of manual testing and testing on real NATs.

Success Criteria

• AutoNAT is fast. That is, I can tell if I'm behind a nat within seconds of starting.
• AutoNAT can't be used as a mild DoS service (can't be tricked into connecting to other IP addresses).
• We can comfortably enable AutoNAT on all DHT servers without worrying about overloading them.
• AutoNAT can reliably determine that a node is dialable (i.e., reduce false positive NATs).

TODO:

• AutoNat and Observed Address Manager Merge AutoNAT with the ObservedAddressManager #900

[willscott]

first wave of functional changes have landed in libp2p/go-libp2p-autonat#41 and libp2p/go-libp2p-autonat-svc#45

Next up is a refactor to make the interface to autonat a bit cleaner, and then cutting a release of that package.

[willscott]

Refactor of the interface and faster responsiveness has landed in libp2p/go-libp2p-autonat#65 / v0.2.3.

Remaining work is primarily around minimizing 'bad addresses', and getting the autonat logic in sync with the ObservedAddressManager. That's tracked in #900