This is similar to CSIT-1789 but that was affecting only AVF plugin, this issue was discovered with DPDK plugin.
There is some possibility VPP code is correct (just CSIT is misconfiguring the tests in some way), but when discussing CSIT-1789 it was agreed RSS behavior in VPP should be deterministic by default.
This issue was noticed in 23.02 release results, but it was already visible in trending, for example here [0].

As usual with RSS issues, they are not visible in most test cases, as they use hundreds of flows (to get fairly even distribution of flows between VPP worker threads). Only some "tunnelling" test with small number of tunnels are sensitive. For testbeds with hyperthreading off, 1-core tests are not affected.
I used [1] a special traffic profile to prove the tunneling does not matter, even l2 cross connect with two (unidirectional) flows can distribute the work differently (compare [3] and [4]).

It is possible other plugins / nics / architectures are also affected (I will add them here when confirmed), 3n-snr was just the first one not still busy with release testing.

[0] https://csit.fd.io/trending/#eNqFkMEOgjAQRL-mXsweqDZw8SDyH6TQFYhQ6rZI8OsFYrJwMF66SWcmb3Z96Alzj-1FqFTEqZBxY-ZHnK7Hebycg5MFbwmkqlC6CBMpyycYZx7QuHMeBmux9RCpKClAloChnv8H48aGsBo0mSjY1o-Lu9AeoSNaAPK2AMwQdjRWXD2x8rMD-zWh5sCmGlsC-g3tX1HO3Ul36Js3cnhdlx3lfEUWZblnhslt1O_-cSZUdrA9devpVfYBz_9y2Q
[1] https://gerrit.fd.io/r/c/csit/+/38381
[2] https://s3-logs.fd.io/vex-yul-rot-jenkins-1/csit-vpp-perf-verify-master-3n-snr/18/log.html.gz#s1-s1-s1-s1-s2-t1-k2-k9-k9-k14-k1-k1-k1-k1
[3] https://s3-logs.fd.io/vex-yul-rot-jenkins-1/csit-vpp-perf-verify-master-3n-snr/19/log.html.gz#s1-s1-s1-s1-s2-t1-k2-k9-k9-k14-k1-k1-k1-k1