Re: [core] Possible false-duplicate detection

[Carsten Bormann <[e-mail]>]

Hi Maciej,

yes, that risk is intrinsic in randomly generating message-IDs.
Section 4.4.2.3 of draft-ietf-lwig-guidance-02.txt contains some recommendations on choosing message-IDs.

If you do need to go random for some reason (e.g., no memory for keeping per-peer counters even for EXCHANGE_LIFETIME), you could randomize the source port (and, possibly, even the source IP address) as well, turning the probability of a false positive from 1e-5 to 1e-9.  By choosing a different token, you can detect the false positive with very high probability, but (for POST) I haven't really thought about recovery strategies that would fully maintain "at-most-once" semantics.
In many cases with very simple clients (which only ever send one kind of request), the damage may just be a lost update (if this is for POST), or simply getting a copy of old data (if this is for GET and the server does buffer the old response, which many don't).

Grüße, Carsten