From my experience managing production web services, raw system metrics only go so far. What’s been more useful is tying metrics to services:
CPU steal time: For cloud VMs, this tells us if noisy neighbors are affecting performance.
Filesystem inode usage: We had a weird outage once where inode exhaustion took down the app.
Service-specific health checks (e.g., latency, request errors): These are often better early indicators than system metrics.
So now we start with “what defines app health?” and trace downward, not monitor all hardware stats and hope to catch something useful.