On Wed, May 30, 2018 at 05:19:26PM -0700, Nick Mathewson wrote:
>In proposal 275, we give reasons for dropping the published-on
>field from consensus documents, to improve the performance of
>consensus diffs. We've already changed Tor (as of 0.2.9.11) to
>allow us to set those fields far in the future -- but
>unfortunately, there is still one use case that requires them:
>relays use the published-on field to tell if they are about to fall
>out of the consensus and need to make new descriptors.
Makes sense. I agree this is worth fixing.
> 2. Mechanism One: The StaleDesc flag
>
>Authorities should begin voting aon a new StaleDesc flag.
>
>When authorities vote, if the most recent published_on date for
>a descriptor has over DESC_IS_STALE_INTERVAL in the past, the
>authorities should vote to give the StaleDesc flag to that relay.
>
>If any relay sees that it has the StaleDesc flag, it should upload
>some time in the first half of the voting interval. (Implementors
>should take care not to re-upload over and over, though: Relays won't
>lose the flag until the next voting interval is reached.)
>
>(Define DESC_IS_STALE_INTERVAL as equal to
>FORCE_REGENERATE_DESCRIPTOR_INTERVAL.)
I think this is the mechanism we should pick.
But I think it needs a name that is more than just a tiny typo away from
Stable. :) Maybe "OldDesc"?
Also, I think you don't mean "the first half of the voting interval".
Maybe you mean the first half hour after that consensus document
appears (not sure what we call that time period), but actually the relay
won't necessarily even *get* the new consensus until partway through
that period, and maybe not even until later depending on what sort of
update schedule it's on. But I agree with the notion that the relay
should try to upload a new descriptor sufficiently before the next vote
starts, and also it should limit the number of times it tries in
reaction to any given consensus it sees.
> 3. Mechanism Two: Uploading more frequently when rejected.
>
>Tor relays should remember the last time at which they uploaded a
>descriptor that was accepted by a majority of dirauths. If this
>time is more than FAST_RETRY_DESCRIPTOR_INTERVAL in the past, we
>mark our descriptor as dirty from
>mark_my_descriptor_dirty_if_too_old().
The reason I prefer the OldDesc mechanism is because trying to keep state
on the relay is going to be hard, given that directory authorities share
descriptors with each other. That is, if you upload your descriptor to
nine dir auths, and eight of them reject it but one of them accepts it,
then it might or might not be the case that all nine of them happily
have the new descriptor by the next voting period -- and if one of them
does get your descriptor from another and decide to like it that way,
then it's not straightforward for the relay to know that that's happened.
(An example of possible pathology: a relay uploads a descriptor D1 to
an authority and it's rejected, and then the authority learns about D1
from another authority's vote and fetches it and likes it, and then the
relay makes a descriptor D2 because it thinks D1 wasn't liked, but then
D2 isn't sufficiently new compared to D1 so the authority rejects D2.)
> 4. Implications for proposal 275
>
>Once most relays are running verions that support the features
>above, and once authorities are generating consensuses with the
>StaleDesc flag, there will no longer be a need to keep the
>published time in consensus documents accurate -- we can start
>setting it to some time in the distant future, per proposal 275.
Sounds like a good plan.
--Roger
___
tor-dev mailing list
tor-dev@lists.torproject.org
https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-dev
