We're using Cacti with RRDTool to monitor and graph about 100,000 counters spread across about 1,000 Linux-based nodes. However, our current setup generally only gives us 5-minute graphs (with some data being minute-based); we often make changes where seeing feedback in "near real time" would be of value. I'd like approximately a week of 5- or 10-second data, a year of 1-minute data, and 5 years of 10-minute data. I have SSD disks and a dual-hexa-core server to spare.
I tried setting up a Graphite/carbon/whisper server, and had about 15 nodes pipe to it, but it only has "average" for the retention function when promoting to older buckets. This is almost useless -- I'd like min, max, average, standard deviation, and perhaps "total sum" and "number of samples" or perhaps "95th percentile" available. The developer claims there's a new back-end "in beta" that allows you to write your own function, but this appears to still only do 1:1 retention (when saving older data, you really want the statistics calculated into many streams from a single input. Also, "in beta" seems a little risky for this installation. If I'm wrong about this assumption, I'd be happy to be shown my error!
I've heard Zabbix recommended, but it puts data into MySQL or some other SQL database. 100,000 counters on a 5 second interval means 20,000 tps, and while I have an SSD, I don't have an 8-way RAID-6 with battery backup cache, which I think I'd need for that to work out :-) Again, if that's actually something that's not a problem, I'd be happy to be shown the error of my ways. Also, can Zabbix do the single data stream -> promote with statistics thing?
Finally, Munin claims to have a new 2.0 coming out "in beta" right now, and it boasts custom retention plans. However, again, it's that "in beta" part -- has anyone used that for real, and at scale? How did it perform, if so?
I'm almost thinking about using a graphing front-end (such as Graphite) and rolling my own retention backend with a simple layer on top of mmap() and some stats. That wouldn't be particularly hard, and would probably perform very well, letting the kernel figure out the balance between frequency of flushing to disk and process operations.
Any other suggestions I should look into? Note: it has to have shown itself able to sustain the kinds of data loads I'm suggesting above; if you can point at the specific implementation you're referencing, so much the better!