Back when we all ran bare metal in colos, you had exactly two tools: whatever the switch exposed via SNMP, and whatever you could scrape over the network. Kids these days think they need a 200MB Go binary on every VM to know if the CPU is warm. You do not.
Option 1: SNMP-only
Enable snmpd, lock it to your poller IP, read the standard OIDs. Load, memory, disk, network. Takes 2MB RAM. Works on a 128MB OpenVZ slice from 2011, mark my words.
Option 2: blackbox_exporter
Run it on one decent box. Probe the rest for HTTP, TCP, ICMP, whatever you exposed. You learn the service is up; you infer the host is not on fire. Good enough for a $5 VM hosting your static blog.
Both approaches leak nothing to the provider's monitoring. Both survive apt upgrade without surprise dependency hell. In twenty years in this business, the simple stuff still wins. The fancy agent will be abandoned by its vendor in eighteen months, and you'll be migrating again.
