Source-based routing across two Linux uplinks
Use separate routing tables so replies leave through the intended provider.
With one default route, a host may answer a connection from the second provider through the first. Source-based policy rules send replies from each public address through the matching uplink.
Example network
eth0 203.0.113.10/29 gateway 203.0.113.9 (ISP-A, primary)
eth1 198.51.100.20/28 gateway 198.51.100.17 (ISP-B)
Routing tables and rules
echo "101 ispa" >> /etc/iproute2/rt_tables
echo "102 ispb" >> /etc/iproute2/rt_tables
ip route add 203.0.113.8/29 dev eth0 src 203.0.113.10 table ispa
ip route add default via 203.0.113.9 table ispa
ip route add 198.51.100.16/28 dev eth1 src 198.51.100.20 table ispb
ip route add default via 198.51.100.17 table ispb
ip rule add from 203.0.113.10 table ispa priority 1000
ip rule add from 198.51.100.20 table ispb priority 1001
Traffic sourced from the address on eth1 uses the ISP-B table. Replies from 198.51.100.20 should leave through that link.
Reverse-path filtering
If incoming traffic is still dropped, check net.ipv4.conf.all.rp_filter. Strict reverse-path filtering checks that the best route back matches the incoming interface. A multi-uplink arrangement can require loose mode.
sysctl -w net.ipv4.conf.all.rp_filter=2
sysctl -w net.ipv4.conf.eth1.rp_filter=2
Persist the configuration
With systemd-networkd, use [RoutingPolicyRule] and a Table= setting in [Route]. With Netplan, define routing-policy and a table for the relevant route.
Run ip route get 1.1.1.1 from 198.51.100.20. The result should include via 198.51.100.17 dev eth1 table ispb.