When conntrack runs out
Every connection through a Linux router with a stateful firewall occupies a conntrack entry. The table is finite, and when it fills, new connections are dropped — often while the host itself looks perfectly healthy.
The symptom
dmesg | grep -i conntrack
# nf_conntrack: table full, dropping packet
The first connections after the table fills are dropped, then some old entries expire and space appears. From the outside it looks like random timeouts on new connections while existing sessions keep working.
What to check before raising the limit
conntrack -Candsysctl net.netfilter.nf_conntrack_max— how full, how big.- Which hosts hold the most entries:
conntrack -L | awk '{print $5}' | cut -d= -f2 | sort | uniq -c | sort -rn | head. - Timeouts: a default
nf_conntrack_tcp_timeout_establishedof five days means every long-lived connection pins an entry.
The usual real cause
A load balancer or monitoring system that opens a new connection per check, with no keepalive. Fixing the client to reuse connections often halves the table without touching a single sysctl.
Only after that: raise nf_conntrack_max, raise the hash table size (nf_conntrack_buckets), and lower timeouts for states you do not need to keep for days.