Misleading Percentiles

[lamont]

I just noticed that zabbix appears to report the percentiles on a given graph after it is done processing it.

So, if you have highly sampled data that you're displaying with 'max' you will get the percentile of the maxes.

I'd consider that almost useless and misleading and never what I'd really want (at least I can't consider of a case where I want that).

What I expected to get was the 95th percentile of the raw data, and then to get a graph of the data with the max values per minute. The 95th percentile of the graphed data was not useful.

I would suggest applying the percentile computation before munging the data for the graphing and not after munging it.

[MartinJanda]

I agree. "True percentile" would be more useful. Alternatively, when loading data after a minute, could use the opportunity to set the percentile to be calculated after 5 minutes, for example.

[erikklaassen]

Is there yet a way to calculate the real 95 percentile?
I have done some calculations based on raw 5min samples, zabbix can be off by 25%! Thats way off.

It seems zabbix calculates the 95% based on the trends table where data is stored every hour. Can this be changed to the history table to get the 95% based on 5min samples?

[erikklaassen]

I now solved this by forcing zabbix to take the history data instead of the trend data.

I now have graphs with a good 95% line, only for graphs with 1 item. If you have two items in you graph line input and output traffic and you want the 95% of both, you have to set your percentile to 97,5 because zabbix collects twice as much data.