[tor-bugs] #17908 [Metrics]: Modify dirauth stats for fallback directories

[Tor Bug Tracker & Wiki]

#17908: Modify dirauth stats for fallback directories
-----------------------------+-----------------
     Reporter:  teor         |      Owner:
         Type:  enhancement  |     Status:  new
     Priority:  Medium       |  Milestone:
    Component:  Metrics      |    Version:
     Severity:  Normal       |   Keywords:
Actual Points:               |  Parent ID:
       Points:               |    Sponsor:
-----------------------------+-----------------
 Currently, the directory authorities estimate client numbers using the
 number of client requests they receive.

 When we introduce fallback directories (likely in 0.2.8, see #17158), the
 directory authorities will receive fewer client queries, because clients
 will choose from the combined set of fallback directories and directory
 authorities when downloading their initial consensus.

 There is a tradeoff here: moving load to the fallback directories,
 decreases the accuracy of client stats, but reduces the load on the
 authorities. I think we could reduce the load on the authorities to 1% of
 the current load, and still get accurate statistics.

 We can still estimate client numbers by multiplying the remaining load on
 the authorities by the proportion of traffic we'd expect the authorities
 to see from a particular Tor version.

 So, for example:
 * Tor 0.2.7 and below: 1x
 * Tor 0.2.8: 100x
   * (Assuming that 1% of initial client directory requests go to the
 authorities)

 This would require us to know the weights used by clients to select
 fallback directories and authorities, implementing #17907 would make this
 much easier.

 For really accurate statistics, we'd also want to know:
 * how many clients only fetch from authorities (FetchDirInfoEarly 1,
 UseFallbackDirs 0)
 * how many relays fetch from authorities (DirPort or dir_server_mode()
 with #12538, Exits, OR relays that don't know their address)
 * how many relays fetch from fallbacks (bridges, OR-only relays,
 hibernating relays)
 But this might not be possible - can we tell relays from clients at the
 moment?

--
Ticket URL: <https://trac.torproject.org/projects/tor/ticket/17908>
Tor Bug Tracker & Wiki <https://trac.torproject.org/>
The Tor Project: anonymity online