it looks like some socket methods can block doing dns resolution?

[njsmith]

I was just investigating some performance issues in trio, and in the process noticed an issue in curio that it looks like anyio might share: anywhere you accept a socket address and pass it through to the stdlib socket module, the socket module will happily accept a hostname and then do a blocking DNS lookup. This is pretty annoying, but whaddaygonnado.

Asyncio and Trio both handle this by automatically resolving hostnames on their own, before passing them to the stdlib socket module. Curio doesn't seem to handle it, and I guess expects the user to pre-resolve all hostnames.

Looking at the v1 code currently, I think anyio's bind method does use a thread when it detects that the address has a hostname in it, but all the other methods that take addresses – connect, sendto, etc. – all pass the address through to stdlib socket directly, so can accidentally block doing DNS lookups.

[agronholm]

You may be right about sendto(), but connect_tcp does a DNS lookup in a thread first, and only passes raw addresses to sock.connect().

[njsmith]

Oh, so sock.connect is a private/internal API?

[agronholm]

Yup. I'm still trying to figure out if asyncio is capable of supporting an async socket abstraction via transports and protocols, but until then, the internal Socket class is private.