Re: `user-mail-address' for message-ids

[Florian Weimer <fw@deneb.enyo.de>]

"Jason R. Mastaler" <jason@libertine.org> writes:

> > As long as the Gnus algorithm provides enough entropy, message IDs
> > generated by Gnus will not cause collisions.
> 
> Can you expand on this a bit further?  In particular, is it feasible
> to arrive at the sufficient amount of entropy with the LHS only?

If only entropy is considered, yes.  If we use 130 bits of entropy
(which would result in 26 characters in base32), and there are 2^40
message IDs to be assigned under a single domain, the probability of
at least one collision among all IDs is around 10^-15 (at least I
think so).  If there 2^50 message IDs, it goes up to 10^-9, if there
are only 2^30, it's approximately 10^-21.  If you add four more
characters (or 20 bits), the probability in the 2^50 case is around
10^-15 again.

If a special domain part is used, only collisions among Gnus users
have to be avoided, so if we assume that the random string in the
domain part is reused for each message once it has been generated, and
there are 2^20 users, with 80 bits of entropy (16 characters), the
probability of a collision is around 10^-12.  Together with the rest
of the domain part and the local-part, we have around 35 characters in
a complete.

On the other hand, If we put all the entropy goes into the local-part,
the domain part is probably rather short (people who have FQDN are
going to use the old scheme, I think), and we come out around 35
characters as well.

So if my calculations are correct, there is no purely technical reason
for putting the entropy in the domain part.  However, it's probably
out of question to obtain 130 bits of entropy per posted message.