Currently full-time data capture only includes internal metrics (with a small number of exceptions). It would be useful to also capture system metrics related to cpu, memory, and storage. For illustrative purposes attached is a POC data capture tool sysmon.py that captures such information on Linux from /proc/stat, /proc/meminfo, and /sys/block/*/stat that has proven useful for problem diagnosis. Captured information includes the following:

/proc/stat
cpu_user
cpu_nice
cpu_system
cpu_idle
cpu_iowait
cpu_irq
cpu_softirq
cpu_steal
cpu_guest
cpu_guest_nice
ctxt
btime
processes
procs_running
procs_blocked
cpus

/proc/meminfo
memtotal
memfree
buffers
cached
swapcached
active
inactive
active anon
inactive anon
active file
inactive file
dirty

/sys/block/*/stat
sd*.reads
sd*.reads_merged
sd*.read_sectors
sd*.read_time_ms
sd*.writes
sd*.writes_merged
sd*.write_sectors
sd*.write_time_ms
sd*.io_in_progress
sd*.io_time_ms
sd*.io_queued_ms

Similar metrics are available through Windows APIs. Where applicable cumulative counters are preferred over instantaneous values because cumulative counters can be sampled at arbitrary time intervals. In general raw system-specific metrics with a minimum of processing are preferred, leaving it to tooling to subsample as needed and compute useful values for display. (An exception might be for example that sectors could be converted to bytes because sector may be a system- or device-specific unit.)