Ian,
You have already identified 3 of the major issues, but I would start by asking the question what will that level of granularity get you and why do you need it?
Is this just information for charts and graphs? Or is this some performance metric that would be measured and possibly alerted on?
If it's the later of those two then you don't gain anything by short polling periods since your realistic response time to an incident is at best 10 minutes. Secondly if it's the later you will most likely end up having to filter out a ton of false alarms which will render the entire point of 1 second polling useless. If it's the former of these two then the false alarms, missing or incorrect data also render the point of that level of detail meaningless.
I was once tasked with creating a monitoring script that would monitor a service on a pool of servers at the 1 second interval. Trust me when I say it was a nightmare and pointless task; although I did learn a great deal from the experience.
My short answer is 1 minute; anything shorter and you will start running into major issues unless you're going to build a daemonized perl script with lots of error checking at which time you're not really talking about a script anymore, now you're building an application.
Travis
