Fix resolution for host names starting with a decimal digit.

[marcioapm]

No description provided.

[marcioapm]

Hi @s-ludwig
We recently upgraded to 0.8.4 and this is biting us hard.
If you merge it, could you also tag a new version as soon as possible?

[s-ludwig]

The fix in this form is unfortunately breaking the other direction (IP addresses are also supposed to work with use_dns == true). But as a quick/correct alternative I'd suggest to keep the condition and simply wrap the if branch in a try {} catch (SocketException) and let it fall through to the else branch if an exception was caught. That would will mean that a transient exception will be thrown for DNS names starting with a digit, but unless the exception is printed (.toString) at least that shouldn't have a measurable performance impact.

I'll tag a (pre) release as soon as this is merged.

[marcioapm]

Well, in that case I think we should use a statically initialized regex to correctly match ip4 and ip6 addresses. It will be more efficient and more correct than the exception allocation and overhead - what do you think?

[s-ludwig]

Not sure about regex because of the compile-time overhead, but yes that would definitely be better. From an effort/gain point of view the exception based approach is hard to beat, though. The exception allocation and stack unwinding overhead should be dwarfed by the actual DNS lookup anyway.

[marcioapm]

How about this test?
Pretty efficient and seems like it could work, based on the facts that : is not a valid name character and a TLD name cannot start with a digit.
See https://stackoverflow.com/questions/9071279/number-in-the-top-level-domain

bool canBeName(const(char)[] host) {
    import std.string : indexOf, lastIndexOf;
    import std.ascii : isDigit;

    if (host.indexOf(':') != -1);
        return false;
    const tld = host.lastIndexOf('.') + 1;
    return !tld || tld >= host.length || !host[tld].isDigit;
}

if (!use_dns || !canBeName(host)) {
  // numeric ...
} else {
  // name ...
}

[marcioapm]

Alternatively we can bypass std.socket, call getaddrinfo directly with AI_NUMERICHOST and interpret the return code, doing the correct thing and avoiding the exception overhead - how do you feel about that?

[s-ludwig]

The canBeName approach looks fine, AFAICS. I'd stay with std.socket if possible, as I was hoping to be able to slowly start reusing parts from Phobos (which is also why scoped InternetAddress instances are used within eventcore instead of a custom struct).

[s-ludwig]

I took the idea and pushed a solution with a slightly modified IPv4 branch: #121