Server-status is decent telemetry but you're still treating the symptom. I've watched Apache workers paint pretty pictures while the whole box suffocates on conntrack table exhaustion.
Contabo's kernel tuning is invisible to you. What does /proc/sys/net/netfilter/nf_conntrack_count look like when these vhosts seize? If you're north of 80% max, those W states are lies the workers tell themselves.
I keep a Grafana dashboard for conntrack on every node I touch. Boring until it isn't.
436 days. reboot is surrender.
