I bet one is S2A and the other S2B, although the dates are not exactly a multiple of 5 days apart (so they are also from different orbits). There is a consistent north-to-south shift between S2A and S2B of about 1 pixel. ESA knows about this, but is apparently waiting for a new processing software upgrade to repair the problem.
More useful if you would provide S2 product details. ReflecKWAK is hard to comprehend.
I understand… This problem difficulties a lot investigation works that need a lot of detail.
Off course, the scene names are:
Repair that the only thing that change in the scenes are S2A instead S2B… And still persist the problem between S2B scenes.
I want to consider one year Sentinel - 2 imagery to study the evolution of that small area, so the problems for me are bigger now.
I agree with the comment that S2A has generally a better geolocation than S2B (especially before January 2018). However it seems from the above image that the shift is very small (lower than one pixel). This kind of shift can be expected for any pair of S2 images. If this is a problem for you than you will need to correct this (for the time being, this will be improved in the future). The shift between two acquisitions of the same tile is generally constant across a tile, except if you look at different repeat orbits over mountainous terrain.