Re: Another data set worth graphing: v3 status votes

[Karsten Loesing <karsten.loesing@xxxxxxx>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hi Roger,

On 04/01/2009 12:42 PM, Roger Dingledine wrote:
> While you're working on metrics, it might be useful to take the
> v3-status-votes file from gabelmoo's datadir (which contains all
> the v3 networkstatus votes for the last consensus), and make a
> graph for each status flag of which authority voted how for each
> relay. I'm imagining relays on the x axis (the votes already sort
> them by identity key I think), and booleans on the y axis, and six
> colors for the six authorities. The votes ought to mostly overlap --
> any significant differences might indicate further bugs. For example,
> if some authorities have stopped voting correctly for Stable, that will
> influence their Guard flag assignment too.

Sure, I made a first attempt to visualize the voting process for some of
the flags. Having all relays on the x axis made the graphs pretty hard
to read, if not impossible. Instead, I arranged the booleans in several
rows, so that all relays having a specific flag fit on one page. This
does not permit comparisons between flags, but maybe that will be step 2
after finding out why single flags behave as they do. Here's the PDF
(2.2 MB):

http://freehaven.net/~karsten/metrics/relayflags-2009-04-01.pdf

> Doing the parsing and graphing in a mostly automated way will let us
> easily rerun the test down the road to make sure that anomalies haven't
> appeared.

The process is fully automated. It's the usual Java-R-LaTeX winning team
that takes only a few seconds to run. (I can upload the sources when we
are happy with the resulting graphs.)

> A more interesting graph might be the average of votes from the authority
> for that relay over some time period. Do we save votes currently, or
> just consensuses?

We save both votes and consensuses. I can have a closer look at the
trends if there is something interesting we hope to find.

> Not super high priority, unless it turns out to uncover a big bug, in
> which case we'll want to go back in time and mark it high priority. ;)

The patterns I found are probably no big bugs, but they do look unusual.
Or they are features that I'm not aware of.

- --Karsten

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFJ1Bny0M+WPffBEmURAtApAJ4yHVVxtikJdAOwJoe9kkLK33nPIwCdEJEA
A0IBwNtNhCXTs/9XfX/45MI=
=4IvE
-----END PGP SIGNATURE-----