Cisco documentation states, that:
If NetFlow is configured for version 7, the flow is performed by the Routing Processor, which could cause high CPU utilization.
For troubleshooting high CPU utilization due to Netflow version 7, configure mls nde sender version 5, as the Netflow export is performed by the SP, which is the default for version 5 or version 9.
It turns out, combination of NetFlow version 9 and NDE sender version 7 also creates high CPU load in certain situations. Here is the setup:
Both routers are Cisco 7604. Other than different IP addresses, the only difference between R1 and R2 was this:
on R1: mls nde sender
on R2: mls nde sender version 5
Default sender version is 7. Both routers configured with ip flow-export version 9.
When ever R2's eBGP session was interrupted, R1's CPU utilization skyrocketed to 100% and stayed there for 10-15 minutes rendering router unusable. "process cpu threshold" reported that "IP Input" was responsible for CPU load, not "BGP Router" as I expected, since these CPU
spikes only happened when eBGP session went down. After changing NDE sender version to 5 on R1, the problem went away.