At Tue, 8 Feb 2011 19:41:41 -0500, Stephen Kent wrote: > > I concur with Geoff re the desirability of the staging period. I > disagree with Rob's suggestion that the staging period is at the > wrong place in the document, or in the phased approach to key > rollover. > > From my perspective, the fundamental rationale for the staring period > is two-fold: > > - we want a delimited interval after which a CA can have > reasonable confidence that all RPs will have retrieved the new CA > key, before transiting to a new CA key and subordinate, signed > products.
I have no problem with -allowing- a CA to wait for a staging interval here. I do still object to -requiring- every CA to wait for a staging interval here. > - we also want a delimited interval for RPs so that they can > acquire a new CA key prior to the transition to that key for > subordinate signed products. Not at all obvious. So long as the new CA key is published before the subordinate signed product, either it just works or you're into the grey area of loose consistency, where there are so many ways to lose that it's almost not worth trying to enumerate them. The net could partition. One of the rsync servers could be down. Your clock could be wrong. Etc ad nausium. > Because of the naming conventions adopted by the repository design, > most of these new products overwrite old ones, so we want to maximize > the likelihood that an RP gets a consistent set of data. You appear to be assuming a model in which an RP discards old good data before fetching new bad data, thus potentially shooting itself in the foot whenever anything changes. Perhaps instead one just ought not to write one's validator that way :) Part of my issue (raised explicitly in an earlier message, not sure if you saw it) is that the justification for the staging period in the draft is very weak, and needs to be rewritten to fix this (if you can) before publication. In essence, it starts with the bad assumption that perfect consistency is possible, even though we know (long mailing list discussion) that it is not, then justifies the staging period by saying that we must maintain consistency for the benefit of hypothetical validators that operate in strange ways that require it. Given that, as far as I know, none of the existing validators do require this of consistency (and given that, as noted, such consistency is not even possible), this does not make much sense. > I agree with Rob that there is no delimited period that will > guarantee that both CAs and RPs are in sync re a CA key transition. > But, that does not mean that we ought not select and publish such a > period, consistent with the other time intervals that we have > published, e.g., the frequency at which RPs SHOULD query the > repository. If I believed in the need for the staging period, I would agree with this. > As a compromise, we could re-phrase the staging period discussion to > RECOMMEND a 24-hour minimum (vs. using MUST here), if the WG (not > just you, Rob, :-)) feels strongly. I would strongly prefer MAY to either MUST or SHOULD here. That said, I would strongly prefer SHOULD to MUST. :) Part of my issue, of course, is that I already have running code which does not follow the model you're trying to mandate here, and I am reluctant to change it to suit an ex post facto specification absent stronger justification than you've provided so far. As far as I have been able to determine, what I have now works just fine, and the changes you're trying to mandate would require not just code changes but some additional complexity in the internal management operations. Right now I have two operations related to key rollover: 1) Generate a new key, ask parent to issue new cert, deprecate (but continue to maintain) old key and cert, and start issuing new product with new key and cert until everything relevant has been reissued. 2) Revoke all deprecated keys and certs. If I accept your changes, I not only need to add a third operation (start using the new key and cert), I also need to make names for the pesky things visible in the control protocol, expose those names in the user interface, etc, because somebody might restart the rollover operation during the staging period. Whole collection of new ways for operator to get confused and shoot self in foot, to no gain that I can see, because you still haven't convinced me that the staging period is necessary in the first place. > I also agree that, as Roque noted, in the case of an emergency rekey, > the 24-hour staging period may be inappropriate. But an emergency > rekey is not the same as a planned key rollover. I suggest that we > note this difference up front and say that, in case of an emergency, > the CA re-issuing the cert which is being re-keyed SHOULD consider > the impact on routing of a quicker transition to a new cert+key, and > behave as it sees fit, given the circumstances surrounding the rekey. So code has to be capable of carrying on without waiting for the staging interval. Good. _______________________________________________ sidr mailing list [email protected] https://www.ietf.org/mailman/listinfo/sidr
