[08:04:37] Krinkle: in theory yes but we had that median on that graph :)ยจ [08:04:43] I'll remove that then. [08:06:03] But I guess we had it to compare with median backed save timing? [08:11:09] I've changed it to not use coal and created a task, maybe 95 was too spiky for savetimings. [09:39:47] https://github.com/spiral/roadrunner [09:40:44] https://roadrunner.dev/assets/benchmark.png [11:38:17] godog: do we have a policy for Grafana plugins? Today we have screenshots/videos on some of our dashboards and I'm looking into adding more meta data to those dashboards by getting a JSON and to do that I need https://grafana.com/grafana/plugins/ryantxu-ajax-panel/ ? [12:12:26] Krinkle: https://docs.google.com/spreadsheets/d/1fca37VmUzQjJwYq6oDRgaoawUmmU9FseUqvQGHfjIzg/edit#gid=1772070593 [12:19:51] I'm making the desktop one now and I'll make maps for PLT as well shortly after [12:21:10] Cool. I'm curious how it will relate [12:28:42] Krinkle: all 4 maps are there now https://docs.google.com/spreadsheets/d/1fca37VmUzQjJwYq6oDRgaoawUmmU9FseUqvQGHfjIzg/edit#gid=2109398186 [12:29:04] phedenskog: not a policy per-se but we do use some plugins, see https://gerrit.wikimedia.org/r/plugins/gitiles/operations/debs/grafana-plugins/+/refs/heads/master [12:39:12] gilles: wow those looks really cool [12:44:34] The difference in mobile cpu/plt between north and south America is really interesting. [14:23:06] gilles: interesting, there seems to be an inverse relation between cpu and plt equity. Meaning that despite cpu inequality, South America still has plt equality? [14:23:32] that's potentially an indicator that the bottleneck is shifting away from cpu and (back) to bandwidth or ram [14:24:06] possibly because the base level cpu has improved so much that inequality no longer matters there for perf [14:24:37] I guess maybe not so surprising actually, this is plt, not mwloadend. so JS cost isn't really there very much [14:25:02] which isn't for granted, we have that because we don't use JS in the critical path [16:26:30] I think in some regions networks sucks across the board and income/device cost becomes a differentiatior [16:27:54] in other words, in some countries money can buy you a faster mobile device but it can't buy you good network [16:28:47] this is the map for income equality https://secureservercdn.net/45.40.146.38/4fe.d3c.myftpupload.com/wp-content/uploads/2014/12/Wonkblog-palma-inequality-map-1024x506.jpg [16:29:30] it looks a lot like the Mobile CPU map I just put together [16:30:38] I can try to apply the same color scheme [16:32:55] done [16:35:09] gilles: I see what you mean, but I thnk we might be seeing the inverse for our sites. If I'm reading this right, we have fairly good PLT equality on mobile in south america, but inequality in CPU. That means the end result is people mostly have the same perceived performance, so the inequality in CPU didn't "help" some. [16:35:30] although the mistake I'm making is that green PLT quality means good performance itself, which is liely not true [16:35:35] yes, their faster device doesn't help [16:36:30] so CPU is indeed still the "new" bottleneck, but only if the network speed is up to a certain base level, which is isn't there. [16:38:07] yep [16:41:07] and some countries have a bit of everything, inequality of network quality based on rural/urban and income inequality that results in people having a bigger gap in terms of high-end and low-end devices [16:43:11] those maps are interesting but I think they bring more questions than they answer. I'm sure they'd be successful if we blogged about them, because people like visual data, but any analysis without further research would be a bit subjective [16:43:32] we're just trying to guess the narrative but we're not confirming it scientifically [16:44:05] it's definitely interesting in itself that they're so different, though [16:44:12] it confirms that the benchmark brings novel signal [16:45:05] if it was tightly coupled to PLT the maps would look the same [16:50:56] yeah [16:52:05] I guess one could potentially confirm a theory of sorts by correlating the country level CPU/PLT evaluation with a per-sample correlation between CPU/PLT. To e.g. confirm that indeed the two are not highly correlated [16:53:04] but it'd be hard to account for other factors, because I'm sure in lots of countries the two would be correlated if done at a global level, merebly because in wealthier countries banwidth, DC latency and CPU are all three "better" [21:13:44] https://infrequently.org/2021/03/the-performance-inequality-gap/