Re: [Analytics] Something's up with EventLogging since Jan 7th

Does this throughput rate match what you see on the database? It's hard to tell, because I can't infer how many events per day are generated based on what you're showing me. It depends on the granularity of the curve's rendering. Even if I knew what the numbers going up to 6.2 on the left represent, it still wouldn't tell me how many per hour/per day are recorded there. Are there a bigger number of events that might not be validating for other > wikis versus China? Are you talking about syntactically incorrect events that don't pass the schema check? Wouldn't this happen if your product is "more used" in china than > elsewhere? Sure, but that doesn't really matter, because global trends would evolve slowly. The point of tracking the network performance statistics worldwide, despite their very uneven distribution, is to immediately spot when something very wrong happens. We also track the same stats per wiki, and per country (albeit not on a timeline at this point for the per-country ones). The bottom line is that the per-country share of recorded EL events shouldn't vary wildly like that. Logs are at: stat1002:/a/eventlogging/ gilles@stat1003:~/20150120$ cat all-events.log-20150120 | grep "\"clientValidated\": true" | grep -c "\"schema\": \"MultimediaViewerNetworkPerformance\"" 330061 And in the DB across the tables that have "active" versions, I get a total of 31 + 327867 = 327898 events recorded A difference of 2000ish events, might not be a real difference since the date cutoff probably differs between the log file and the timestamp contained in its entries. Now, looking at an earlier date, when things seemed to be stable in terms of country-based sampling: gilles@stat1003:~/20150103$ cat all-events.log-20150103 | grep "\"clientValidated\": true" | grep -c "\"schema\": \"MultimediaViewerNetworkPerformance\"" 287712 And in the DB, a total of 5 + 289461 = 289466 So far it does look like everything is making it to the DB, I'll keep investigating tomorrow. On Wed, Jan 28, 2015 at 5:43 PM, Nuria Ruiz <nuria(a)wikimedia.org> wrote: > Gilles: > This event has a "pretty" constant rate of input: > http://graphite.wikimedia.org/render/?width=588&height=311&_salt=14… > And as far as I can see it has not changed (greatly) before and after the > 7th. Does this throughput rate match what you see on the database? If so, I > would say db matches incoming stream. Now, if it doesn't (somehow the db > has less data) we might have a problem. However, dropped events when > inserting on the db will affect all countries equally so I doubt these will > be the source of the discrepancy. > Are there a bigger number of events that might not be validating for > other wikis versus China? > Logs are at: stat1002:/a/eventlogging/archive and the all-events log will > have every single one of your events received. Please note that events are > not inserted right away when received, there is a buffer of couple minutes. > >This balance shifting over time is really problematic for tracking > Media Viewer client-side network performance, because Chinese >traffic > suddenly accounting for a bigger or smaller share of the overall recorded > events creates big swings in the global >averages/percentiles (since > network performance in China is bad). > Wouldn't this happen if your product is "more used" in china than > elsewhere? If you are counting "absolute" values you will be always skewed > to the biggest dataset unless you are doing some kind of scaling. > Thanks, > Nuria > On Wed, Jan 28, 2015 at 4:09 PM, Gilles Dubuc <gilles(a)wikimedia.org> wrote: >> Hi all, >>> I've tracked down an unexplained EL phenomenon that surfaced in our >> stats as a false trend in our global stats. >>> The data I'm looking at specifically is coming from Media Viewer's >> MultimediaViewerNetworkPerformance_* tables. >>> Have a look at this graph: >> https://docs.google.com/spreadsheets/d/1PJsyzAyj74dctGCl4-09L7LS4AMZRh57G56… >> the big change is on Jan 7th/8th >>> It shows how many EL events we've recorded, per client-reported country, >> over the last 90 days. The sampling factor we use has been constant for >> each wiki over that period. Thus, the distribution shouldn't evolve >> drastically, aside from seasonal/local trends. Besides the Ukraine spike on >> a particular date (probably related to world events), the graph before Jan >> 7th looks like what you would expect. Then, following the outage that >> happened on Jan 7th, not only the balance is completely changed, but it >> evolves over time (the US and China are keeping "higher than normal" >> levels, while the rest seems to slide down lower than pre-7th quantities), >> showing me that something strange is happening and is probably unresolved. >>> This balance shifting over time is really problematic for tracking Media >> Viewer client-side network performance, because Chinese traffic suddenly >> accounting for a bigger or smaller share of the overall recorded events >> creates big swings in the global averages/percentiles (since network >> performance in China is bad). >>> _______________________________________________ >> Analytics mailing list >> Analytics(a)lists.wikimedia.org >> https://lists.wikimedia.org/mailman/listinfo/analytics >>> _______________________________________________ > Analytics mailing list > Analytics(a)lists.wikimedia.org > https://lists.wikimedia.org/mailman/listinfo/analytics