Thanks for your explanation.
This is very interesting,especially the 2 scenarios (which low-level metrics would look the same.) I wonder how people are doing today for big web sites, such as Amazon, twitter, flick, youtube? Do they monitoring the application-level metrics? That would be a lot to look at.
Secondly, is web service widely used today for database interfaces, especially in multi-tiered web application systems? Thanks!

Yes. For example, today you often don’t realize that you have a connection leak (connections that are not used and yet not returned to the pool either) until you run into stuck threads or equivalent lack of service. You shouldn’t have to go down to these such low-level concepts for this, it should appear through application-level metrics.

In another example, you often see people having issues (response time too high, connections refused…) and it is not clear at first whether something is going wrong (there is technical problem) or you simply are getting a lot more traffic than usual traffic (nothing broken, just high demand). Because through the low-level metrics (e.g. CPU usage, memory utilization) both scenarios may look the same. A monitoring system that is centered on application-level metrics would make it instantly clear which situation you’re in.

What does “in ways that are related to what the application really does (as opposed to generic metrics like memory, CPU and I/O metrics…)” mean? Do you mean to troubleshoot the applications according to its behavior?