This seems similar to the the problem I brought up a while ago.
I still have not overcome the issue. Actually, recent attempts make me wonder, if it has not got worse.
For me, one back-geocoded band got lost if S1A was a master, but S1B was a slave (always the same subswath, same burst from different slices), I started processing this spring data (only S1A yet) and I already got an empty band for that burst and not always (e.g. data from four dates D1, D2, D3, D4, coherence of D2 and D3 was bad, other two were ok).