Hi David, I can see the point with random constants, even though I think the risk for collisions is pretty small in practice -- especially in controlled experiments. What is more important is what Christian Huitema discusses in another message: to change the constants when the protocol evolves, at least when the spec is stable enough that you reach interop testing.
The reason I mention the multipath draft is of course that it is a wg document that has gone through a few revisions. If you think the constants are important, you should probably start there and not when someone presents a new idea for the first time for discussion. BTW, the One True Way to generate random numbers is to hex dump /dev/urandom :-). /me ________________________________ From: David Schinazi <[email protected]><mailto:[email protected]> To: Michael Eriksson <[email protected]><mailto:[email protected]> Cc: Martin Thomson <[email protected]><mailto:[email protected]>, IETF QUIC WG <[email protected]><mailto:[email protected]> Subject: Fwd: New Version Notification for draft-piraux-quic-additional-addresses-00.txt Date: Tue, 18 Oct 2022 20:28:20 +0200 (Central European Summer Time) Hi Michael, While MT's comment might sound nitpicky, he's right in suggesting people use real PNRGs because we've already had collisions due to human-picked numbers in the past. The fact that quic-multipath made the same mistake doesn't make it best practice. (And FWIW I'm also guilty of having made that mistake in the past). David On Tue, Oct 18, 2022 at 2:28 AM Michael Eriksson <[email protected]<mailto:[email protected]>> wrote: On Tue, Oct 18, 2022 at 11:13:24 +1100, Martin Thomson wrote: > I see this in the draft: > > "TBD - experiments use 0xadda" > > I find it hard to believe that this value was chosen at random. > Please consult a random number generator for these values. And - > while you are developing proposals - larger values might be more > appropriate. That was a pretty nitpicky comment... Have you read draft-ietf-quic-multipath? The 0xbabaXX constants don't look very random if you consider the affiliation of the first authors. Also, what is a "large" value? 0xadda is big enough to require a 32-bit VarInt. /me
