This is different from later regressions related to l2-input and dpdk version.
The regression is detected only for MRR trending, I assume PDR does not have enough samples to distinguish this from other regressions.
Still, This has to be the reason why l2basemaclrn has worse performance comparison than l2xcbase, which has both the l2-input and dpdk regressions, but not this one.

In general, this issue is affecting all platforms, except maybe taishan [0], but even there the spread of results changes, and zen2 [1], where the regression is smaller than usual (so the anomaly detection does not pick it up from the 4 samples available.
Another notable platform is Denverton [2], where this issue also affects l2xcbase.
Except Denverton, this issue is best visible on Skylake [3], mostly by the virtue of reducing the spread.

[0] https://s3-docs.fd.io/csit/master/trending/trending/l2-3n-tsh-x520.html
[1] https://s3-docs.fd.io/csit/master/trending/trending/l2-2n-zn2-xxv710.html
[2] https://s3-docs.fd.io/csit/master/trending/trending/l2-2n-dnv-x553.html#t1c
[3] https://s3-docs.fd.io/csit/master/trending/trending/l2-2n-skx-xxv710.html#t1c