We cannot guarantee that with 60 events a sec things will still work well (as I said we should schedule some perf testing on this regard so I filed an item for this purpose: https://phabricator.wikimedia.org/T86244)
See that now we go beyond 300 events per sec here and there: http://ibin.co/1nTsNYc1bekd I recommend sampling those events 1:10. Thanks, Nuria On Thu, Jan 8, 2015 at 12:06 PM, Ryan Kaldari <[email protected]> wrote: > After talking with Dario and Leila we decided that we will sample the > page-impression event at 1:1000. We would, however, like to retain the > widget-impression event unsampled if possible. That event happens > approximately 50% as often as page-impression. So we're probably talking > about somewhere around 60 events per second in that case. Would that be > acceptable or should we sample the widget-impression event as well? > > Kaldari > > On Wed, Jan 7, 2015 at 5:33 PM, Leila Zia <[email protected]> wrote: > >> Thanks, Nuria! >> >> On Wed, Jan 7, 2015 at 5:30 PM, Ryan Kaldari <[email protected]> >> wrote: >> >>> Thanks everyone for the research on this! I'll go ahead and create a >>> card for implementing sampling on the high-throughput WikiGrok events. >>> >>> Kaldari >>> >>> On Wed, Jan 7, 2015 at 5:20 PM, Nuria Ruiz <[email protected]> wrote: >>> >>>> Sorry, I send it too soon, trying again: >>>> >>>> >We're talking about a total of ~170 events per sec for these pages. >>>> This is to high to log in 1:1 rate, we would need to do 1:10. At this >>>> time most events on EL logging log at a much lower rate, events over 1 per >>>> sec are the following, as you can see mobile & media viewer are the >>>> majority of the throughput. >>>> >>>> My preference would be to be less than 400 events per sec until we have >>>> done some perf testing to make sure we can handle it (we might be able to >>>> as we have done many improvements since we set these thresholds) >>>> >>>> MobileWebClickTracking 41.35% (114.15/sec) >>>> MediaViewer 21.66% (59.78/sec) >>>> MobileWikiAppToCInteraction 12.44% (34.35/sec) >>>> PageContentSaveComplete 3.39% (9.35/sec) >>>> EchoInteraction 2.69% (7.42/sec) >>>> NavigationTiming 2.51% (6.93/sec) >>>> MultimediaViewerNetworkPerformance 1.84% (5.07/sec) >>>> SaveTiming 1.58% (4.37/sec) >>>> Edit 1.39% (3.83/sec) >>>> PersonalBar 1.24% (3.43/sec) >>>> TimingData 0.83% (2.28/sec) >>>> MobileWebUIClickTracking 0.73% (2.02/sec) >>>> Popups 0.68% (1.87/sec) >>>> MobileWikiAppOnboarding 0.62% (1.70/sec) >>>> MultimediaViewerDimensions 0.61% (1.68/sec) >>>> UniversalLanguageSelector 0.50% (1.37/sec) >>>> PageCreation 0.50% (1.37/sec) >>>> MultimediaViewerDuration 0.47% (1.30/sec) >>>> MobileWebEditing 0.45% (1.25/sec) >>>> MobileWikiAppSearch 0.41% (1.13/sec) >>>> CentralAuth 0.40% (1.12/sec) >>>> >>>> On Wed, Jan 7, 2015 at 5:12 PM, Nuria Ruiz <[email protected]> wrote: >>>> >>>>> >We're talking about a total of ~170 events per sec for these pages. >>>>> This is to high to log in 1:1 rate, we would need to do 1:10. >>>>> >>>>> On Wed, Jan 7, 2015 at 4:10 PM, Leila Zia <[email protected]> wrote: >>>>> >>>>>> Thanks everyone for chiming in. Your comments were very helpful. :-) >>>>>> >>>>>> Nuria, I checked the per second pageview count for the pages wikigrok >>>>>> will be live on for 3 hours in 2015-01-07 (as a sample). We're talking >>>>>> about a total of ~170 events per sec for these pages. Of course major >>>>>> events can affect this number. This number added to the current 270 >>>>>> events >>>>>> per sec you mentioned will send us over the 350 events per sec limit (if >>>>>> it's a hard limit). What do you think? >>>>>> >>>>>> Leila >>>>>> >>>>>> >>>>>> >>>>>> On Wed, Jan 7, 2015 at 10:13 AM, Nuria Ruiz <[email protected]> >>>>>> wrote: >>>>>> >>>>>>> >Given that information, do you have any idea if we are in danger >>>>>>> of overloading EventLogging? >>>>>>> Logging broad events (such a page load) 1 to 1 might incur into >>>>>>> problems as our traffic is high enough that events logged1/1000 happen >>>>>>> still in very large amounts. >>>>>>> >>>>>>> Some numbers (oversimplyfying and rounding) >>>>>>> >>>>>>> We have about 200 million visits per day for the enwiki mobile site >>>>>>> . This means about 2300 pageviews per sec, if we are sending 1 load >>>>>>> event >>>>>>> per pageview EL will (sadly) die, most likely. >>>>>>> >>>>>>> If we assume EL handles up to 350 events per second (and now we are >>>>>>> at 270 events per sec) I would think that sending 10 events per sec on >>>>>>> your >>>>>>> case would be pretty safe. That would be sampling about 1/200 for a load >>>>>>> event per every pageview. This seems like a good upper bound. >>>>>>> >>>>>>> Now, since there are no constrains as to how long you keep your >>>>>>> experiment running you can try a lower sampling ratio, say, 1/1000 and >>>>>>> keep >>>>>>> the experiment running for longer. >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Tue, Jan 6, 2015 at 5:50 PM, Ryan Kaldari <[email protected] >>>>>>> > wrote: >>>>>>> >>>>>>>> The highest volume events we are going to log will be: >>>>>>>> 1. For each of the 166,000 articles, one event when the page loads >>>>>>>> 2. For each of the 166,000 articles, one event when the WikiGrok >>>>>>>> widget enters the viewport (about half as often as #1) >>>>>>>> >>>>>>>> These will be active for all mobile users, logged in and logged >>>>>>>> out, including many high pageview articles. >>>>>>>> >>>>>>>> Given that information, do you have any idea if we are in danger of >>>>>>>> overloading EventLogging? If so, do you have recommendations on >>>>>>>> sampling? >>>>>>>> So far, everyone has said not to worry about it, but it would be good >>>>>>>> to >>>>>>>> get a sanity check for this test specifically. >>>>>>>> >>>>>>>> Kaldari >>>>>>>> >>>>>>>> On Tue, Jan 6, 2015 at 4:57 PM, Nuria Ruiz <[email protected]> >>>>>>>> wrote: >>>>>>>> >>>>>>>>> (cc-ing mobile-tech) >>>>>>>>> >>>>>>>>> Since we do not the details of how wikigrok is used and its >>>>>>>>> throughput of requests we can not "estimate" sampling ourselves. I >>>>>>>>> imagine >>>>>>>>> wikigrok is been deployed to a number of users and it is with that >>>>>>>>> usage >>>>>>>>> the mobile team could estimate the total throughput expected, with >>>>>>>>> this >>>>>>>>> throughput we can recommend sampling ratios. >>>>>>>>> >>>>>>>>> >>>>>>>>> Thanks for asking about this without before deploying! >>>>>>>>> >>>>>>>>> >>>>>>>>> On Tue, Jan 6, 2015 at 4:55 PM, Ryan Kaldari < >>>>>>>>> [email protected]> wrote: >>>>>>>>> >>>>>>>>>> I can elaborate on this after I finished the SWAT deployment.... >>>>>>>>>> Gimme 30 minutes or so. >>>>>>>>>> >>>>>>>>>> On Tue, Jan 6, 2015 at 4:51 PM, Leila Zia <[email protected]> >>>>>>>>>> wrote: >>>>>>>>>> >>>>>>>>>>> Hi, >>>>>>>>>>> >>>>>>>>>>> The mobile team is planning to switch WikiGrok on for >>>>>>>>>>> non-logged in users next week (2014-01-12). The widget will be on on >>>>>>>>>>> 166,029 article pages in enwiki. There are two EventLogging schema >>>>>>>>>>> that may >>>>>>>>>>> collect data heavily and we want to make sure EL can handle the >>>>>>>>>>> influx of >>>>>>>>>>> data. >>>>>>>>>>> >>>>>>>>>>> The two schema collecting data are: >>>>>>>>>>> https://meta.wikimedia.org/wiki/Schema:MobileWebWikiGrok >>>>>>>>>>> https://meta.wikimedia.org/wiki/Schema:MobileWebWikiGrokError >>>>>>>>>>> and the list of pages affected is in: >>>>>>>>>>> wgq_page in enwiki.wikigrok_questions. >>>>>>>>>>> >>>>>>>>>>> It would be great if someone from the dev side let us know >>>>>>>>>>> whether we will need sampling. >>>>>>>>>>> >>>>>>>>>>> Thanks, >>>>>>>>>>> Leila >>>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> _______________________________________________ >>>>>>>>>> Analytics mailing list >>>>>>>>>> [email protected] >>>>>>>>>> https://lists.wikimedia.org/mailman/listinfo/analytics >>>>>>>>>> >>>>>>>>>> >>>>>>>>> >>>>>>>>> _______________________________________________ >>>>>>>>> Analytics mailing list >>>>>>>>> [email protected] >>>>>>>>> https://lists.wikimedia.org/mailman/listinfo/analytics >>>>>>>>> >>>>>>>>> >>>>>>>> >>>>>>>> _______________________________________________ >>>>>>>> Analytics mailing list >>>>>>>> [email protected] >>>>>>>> https://lists.wikimedia.org/mailman/listinfo/analytics >>>>>>>> >>>>>>>> >>>>>>> >>>>>>> _______________________________________________ >>>>>>> Analytics mailing list >>>>>>> [email protected] >>>>>>> https://lists.wikimedia.org/mailman/listinfo/analytics >>>>>>> >>>>>>> >>>>>> >>>>>> _______________________________________________ >>>>>> Analytics mailing list >>>>>> [email protected] >>>>>> https://lists.wikimedia.org/mailman/listinfo/analytics >>>>>> >>>>>> >>>>> >>>> >>>> _______________________________________________ >>>> Analytics mailing list >>>> [email protected] >>>> https://lists.wikimedia.org/mailman/listinfo/analytics >>>> >>>> >>> >>> _______________________________________________ >>> Analytics mailing list >>> [email protected] >>> https://lists.wikimedia.org/mailman/listinfo/analytics >>> >>> >> >> _______________________________________________ >> Analytics mailing list >> [email protected] >> https://lists.wikimedia.org/mailman/listinfo/analytics >> >> > > _______________________________________________ > Analytics mailing list > [email protected] > https://lists.wikimedia.org/mailman/listinfo/analytics > >
_______________________________________________ Analytics mailing list [email protected] https://lists.wikimedia.org/mailman/listinfo/analytics
