[08:45:03] <dcausse>	 ryankemper: nice!
[10:44:16] <dcausse>	 errand+lunch
[11:22:13] * gmodena waves
[13:17:05] <dcausse>	 gmodena: welcome! :)
[13:26:26] <dcausse>	 ryankemper: was doing some checks on the new hosts and I believe they're consuming from the wrong (codfw|eqiad).rdf-streaming-updater.mutation topic, they should be consuming from their respective subgraph topic (https://gerrit.wikimedia.org/r/c/operations/puppet/+/1100812)
[13:27:02] <dcausse>	 sadly this means we need to do another data-transfer because the journal got already "corrupted" by consuming from the "full" graph topic
[14:20:38] <inflatador>	 <o/
[14:37:34] <dcausse>	 o/
[15:02:03] <ebernhardson>	 \o
[15:02:48] <dcausse>	 o/
[15:07:44] <inflatador>	 \moti wave2
[15:07:48] <inflatador>	 .o/
[15:09:16] <inflatador>	 dcausse just merged https://gerrit.wikimedia.org/r/c/operations/puppet/+/1100812 , starting on the data xfers shortly
[15:13:06] <dcausse>	 inflatador: thanks!
[16:53:53] <ebernhardson>	 randomly interesting, there are a surprising number of sessions that click in reverse.   Of the 10 random sessions here, 3 of them first clicked on lower results, then later clicked on the top result
[16:54:27] <ebernhardson>	 i suppose it's a sample of 10 sessions that clicked at least twice
[17:22:42] <dcausse>	 weird...
[17:25:11] <ebernhardson>	 i wonder where the timestamps are assigned...maybe its events delivered out of order
[17:27:33] <dcausse>	 if all of them are in that weird order yes possibly
[17:28:07] <dcausse>	 oh you said 3 out of 10, seems high tho
[17:31:01] <dcausse>	 meta.dt should be assigned by eventgate
[17:38:15] <ebernhardson>	 hmm, i'm using the top level dt. lemme see if there is any difference (iirc there was before, but i think that was squashed)
[17:44:56] <dcausse>	 top level dt should be set by the client but unsure if that stream is following recent conventions
[17:45:08] <dcausse>	 dinner
[17:46:04] <ebernhardson>	 nope, same. Looking at raw event the three seems totally reasonable. If anything the surprising part is that there are 10-15s between click and visit events.  In the overall dataset these make up 1.4%, rather than the 30% suggested
[17:46:17] <ebernhardson>	 maybe simply coincidence in the sample
[17:47:31] <ebernhardson>	 in 1.4% the first clicked position is further down the page than the last clicked. I suppose that includes single clicks though. Of sessions with two clicks its 27%
[17:47:37] <ebernhardson>	 so the 30% was right :P
[17:47:47] <ebernhardson>	 it's only a single hour of enwiki though
[17:48:14] <ebernhardson>	 263 sessions w/ 2+ clicks
[18:12:20] <dcausse>	 still seems high... but I that's just me assuming that you can't miss a good result at the top, can't think of a good explanation other than the snippet was really that attractive, a misclick, or some sort of ui glitches
[18:19:50] <ebernhardson>	 maybe the results are that bad that they dismissed the first result, scanned down, chose something, came back and gave the first result a shot
[19:17:57] <dcausse>	 ah yes seems like a sensible explanation, would have to see few examples
[19:40:55] <inflatador>	 data xfers finished on internal graph split hosts wdqs1026,1027, 2018. 2019 in progress; 2020 (the last one) will start after that
[19:41:58] <inflatador>	 we also need to bring wdqs1025 online
[19:53:50] <inflatador>	 appointment, back in ~90m
[20:38:03] * ebernhardson is slightly disapointed about how much better the old AB test graphs looks, i'm not going to go around finding different styles but the default pyplot is pretty meh :P
[21:51:00] <ebernhardson>	 not really even like one thing, just every little piece is slightly worse in mine :P https://phabricator.wikimedia.org/F57782801
[22:21:56] <Trey314159>	 ebernhardson: While I appreciate a pretty graph more than most, I'll take homely but real and readable data over hypothetical beautiful data any day. Looks good (enough)!