Re: Fwd: New Version Notification for draft-wouters-edns-tcp-keepalive-00.txt (fwd)
Joe Abley <[email protected]>
| Newsgroups | gmane.ietf.dnsext |
|---|---|
| Message-ID | <[email protected]> |
On 2013-10-17, at 06:00, Tony Finch <[email protected]> wrote: > Paul Wouters <[email protected]> wrote: >> >> Historically, DNS transactions over TCP only handle 1 query/answer. > > Really? I was using persistent TCP connections back in the 1990s. They > work really nicely. I think the problem is not necessarily that serious authority server implementations don't implement TCP; the problem is more that TCP support is provided as a fallback mechanism with short server-side timeouts and a small capacity for active sessions, and TCP from clients is generally treated as the exception rather than the rule (we hope they'll use it if we set TC=1, but otherwise we don't really expect it). The goal of this option is to define a signalling mechanism whereby a client and server can both be confident that the other side has set these expectations aside, and instead might well prefer to do TCP over UDP. This has implications for both the server and the client with respect to the assumptions in the previous paragraph, the most obvious of which is for a server to reign in any aggressive close-after-response. The protocol elements in this -00 could surely use improvement. However, the basic idea, as described above, is necessary if we want to find a way to use pools of TCP connections between busy endpoints as an alternative to the usual source-spoofed UDP spray. > Note also that RFC 5966 says that servers MUST support TCP and support > multiple queries per connection. I think edns-tcp-keepalive should be > compatible with that requirement. No argument there. Joe _______________________________________________ dnsext mailing list [email protected] https://www.ietf.org/mailman/listinfo/dnsext