Incident Replay · Networking
When Pods Can't Reach RDS Across a Transit Gateway
Flip the node and network settings, watch the database's reply reach checkout or die, then replay the hang from timeout to fix.
Asymmetric routing in 60 seconds
- Reverse-path check (rp_filter): in strict mode (1), Linux drops a reply arriving on a different network card than it would answer from. 0 is off.
- The trap: custom networking always puts Pods on a second card (eth1). The database's reply lands on eth0, so strict mode drops it; checkout hangs.
- The fix: EKS's network plugin sets eth0 to loose (2: any route back will do); the hardened node image forced strict. Bake loose in.
- Follow the reply (SYN-ACK), not the opening SYN. Flow Logs show the database answered; tcpdump (packet capture) and dmesg (kernel log) show where it died.
- Two other traps: skipping SNAT (the address swap) needs a route back to Pods; a firewall needs appliance mode (one firewall both ways).