UDN work & packet tracing might be a good opportunity to improve performances, by replacing our user-space deduplication with in-kernel deduplication.

NETOBSERV-1996 introduces in-kernel deduplication: all flows for a given 5-tuple are now deduplicated in kernel.

This task is about going one step further and use packet tracing across observation points (interfaces) to identify when a packet was already seen, and find the correlated flow without creating a new one. In this way, NATed packets, or packets on br-ex, which have different 5-tuples (e.g. they have Service IP or Node IP instead of pods IPs) can still be correlated with their corresponding pod-to-pod flow and be deduplicated.