Server monitoring should combine CPU, memory, disk, network, and application signals. A single high load-average value does not prove that the CPU is overloaded, and low CPU usage does not prove that a server is healthy. The goal is to identify a measurable bottleneck and connect it to user-visible symptoms.
Establish a baseline
Capture normal behavior during quiet and busy periods. Useful commands include:
uptime
top
free -h
vmstat 1
iostat -xz 1
df -h
ss -s
Install tools such as sysstat only through the distribution's supported package manager. Compare current values with a baseline rather than reacting to one sample.
Interpret the signals together
Load average represents runnable and uninterruptible tasks over time; it is not a CPU percentage. High CPU with sustained run queues can indicate compute pressure. High I/O wait or elevated disk latency may point to storage contention. Memory pressure may cause swapping, reclaim activity, or out-of-memory kills. A filesystem approaching capacity can break uploads, sessions, logs, and database writes even when CPU and RAM look normal.
For PHP sites, correlate system metrics with PHP-FPM active and queued workers, slow requests, database query time, and web-server response codes. A pool at its worker limit may reflect slow application requests or database contention rather than insufficient hardware.
Choose useful alerts
Alert on sustained conditions and user impact: disk space thresholds, repeated service restarts, high error rates, long response times, failed backups, and database unavailability. Avoid alerts that fire on every short-lived spike. Include context such as host, service, time window, and recent deployment identifier.
Capacity decisions
Before adding CPU or RAM, identify whether the workload is CPU-bound, memory-bound, I/O-bound, or waiting on a dependency. Optimize slow queries, reduce unnecessary background concurrency, and cache suitable read-heavy operations. Record changes and compare measurements afterward. Monitoring is most valuable when it leads to a testable hypothesis rather than a reflexive server upgrade.