Consider collectd (http://collectd.org/). Unlike a lot of the usual suspects, collectd is a daemon that records the usual server health stats every 10 seconds into rrd files. After running it for two years on a few dozen systems, it's never failed or caused undue load on its own. We often see events that would have gone completely unnoticed in a 5-minute monitoring window.
It looks like complete overkill for a single server and I don't know how useful it is but it does look promising. It certainly looks like it would scale to tens if not hundreds of machines easily.
It includes an alert structure so different events triggers different actions. For example if the database stops responding email the DBA, if it's a router email the network admin etc.
Again I can't vouch for it over the long term as it's only been a week or so of testing but I can't complain atm.
It's a PITA field to research and I'm trying to avoid the 'roll my own' urges as I'd quite like to write it ;)
Anyone else got any ideas? There is a python based monitoring application out there somewhere that I stumbled upon about 6 months ago with a great plugin API and neato graphs but I can't find it again :( I blame google and not my incompetence :)
Nagios is excellent once you get comfortable with it. People have written check scripts for all sorts of bizarre hardware, interfaces (RS-232 polling, X10, etc.), and so forth. You can turn the babysitting proactive by writing event scripts, eg. when a MySQL slave goes out of sync or memcached wedges, the event handler notices, kicks it, and if it doesn't recover after a few tries, THEN you get a page. Obviously you need to be careful about what you put in an event handler, but used judiciously, they're great.
Cacti, ganglia, and so forth are useful adjuncts for routers, clusters, and the like. But Nagios is a time-tested warhorse with a lot of community support. It's ugly but it works and works well.
PXE + cfengine + nagios can be pulled together for Real Ultimate Power, or at least a simulacrum of what goes on at places like Google (whose scripts are custom Python monstrosities for the most part, but the functionality is pretty similar, with several levels and types of babysitter systems funneling into a more proactive statistical resource allocation type of analysis framework since the scale is so vast; at least that's the track that development was on when I left, and Urs is still in charge so I doubt it's swerved much).
See here for an article on adaptive state monitoring: http://www.onlamp.com/pub/a/onlamp/2006/05/25/self-healing-n... with Nagios and Cfengine. Add something like PXE reimaging of dead nodes to the mix and you cut down workloads by an order of magnitude on large installations.
If you are a sysadmin and haven't experimented with self-healing systems, you should fix that gap in your skill set. If you have a sysadmin who can't or won't, fire him.
If you're starting from scratch, consider puppet (http://www.reductivelabs.com/projects/puppet/) as a cfengine replacement. cfengine is solid, but unless you've already mastered it, I think you'll find puppet is at least as capable and a lot more intuitive.
If you want to measure the load on a server, look into the sar utility, which is included in Linux and all other Unix'es. It will take snapshots of data at 10 minute intervals and store them. Other tools will take that data and turn it into a bar graph or other visualization.
Rock Solid Arcade I suspect at many points is outgrowing my cacheless hit the database on every page setup. I started out with a cron job monitoring top every minute, and many times the usage hits 50% - I'm using Django, so the GIL means that's at full capacity, but I have no idea how instantaneous that figure is.
I got some helpful advice on the Django chat room to monitor Apache, but really what I'd like is for some warning to be dumped to a log if a connection queue backs up for more than a second (i.e. the server is at full capacity). That'll be time to cache/upgrade. :)
I used nagios at the last place I worked with a number of servers, and I was pretty happy with it, although I could very easily imagine something better.
