Multiple left joins or aggregations in one query produce incorrect results

Structured streaming queries which involve more than one non-map-like transformations (either left joins or aggregations) produce incorrect (and sometimes inconsistent) results.

I have put up a sample here as Gist https://gist.github.com/jammann/b58bfbe0f4374b89ecea63c1e32c8f17

The test case involves 4 topics, and I've implement 6 different use cases of which 3 fail to produce the expected results

1) A innerjoin B: to topic A_B
expected: 10 inner joined messages
observation: OK

2) A innerjoin B outerjoin C: to topic A_B_outer_C
expected: 9 inner joined messages, last one in watermark
observation: OK

3) A innerjoin B outerjoin C outerjoin D: to topic A_B_outer_C_outer_D
expected: 9 inner/outer joined messages, 3 match C, 1 match D, last one in watermark
observation: NOK, only 3 messages are produced, where the C outer join matches

4) B outerjoin C: to topic B_outer_C
expected: 9 outer joined messages, 3 match C, 1 match D, last one in watermark
observation: NOK, same as 3

5) A innerjoin B agg on field of A: to topic A_inner_B_agg
expected: 3 groups with a total of 9 inner joined messages, last one in watermark
observation: OK

6) A innerjoin B outerjoin C agg on field of A: to topic A_inner_B_outer_C_agg
expected: 3 groups with a total of 9 inner joined messages, last one in watermark
observation: NOK, 3 groups with 1 message each, where C outer join matches