I noticed that a Joiner doing an inner join is often much faster if prepended with a Reference Row Filter. Both nodes combined beat the single one by far. Of course now I can do that manually, but I can't think of a reason why this optimization should have negative consequences inside the Joiner itself, if done carefully. Or maybe it actually is integrated, but too carefully?
Of course there's a bigger story behind this: This pattern often arises when re-adding data that is for some reason "necessarily" lost in a branch. In these cases the Joiner becomes kind of a multi-column Cell Replacer - but actually this use case falls exactly in a gap between the current capabilities of both. So adding multi-column-capability to the Cell Replacer would be another way to solve this issue - but then there would be two competely parallel ways to do an inner join...