Gustavo de Morais created FLINK-40831:
-----------------------------------------
Summary: Mini-batch outer join loses the null-padded row when an
input record is updated without UPDATE_BEFORE
Key: FLINK-40831
URL: https://issues.apache.org/jira/browse/FLINK-40831
Project: Flink
Issue Type: Sub-task
Components: Table SQL / API
Reporter: Gustavo de Morais
Assignee: Gustavo de Morais
{{MiniBatchStreamingJoinOperator}} shares the per-record logic of
{{{}StreamingJoinOperator{}}}, so it has the same problem. When a join input
has no UPDATE_BEFORE, an UPDATE_AFTER can replace a stored record with the same
unique key, and the operator counts it as a new match on the outer side. After
the record is deleted, the count doesn't reach 0, and the null-padded row is
never emitted again.
Mini-batch adds one more case. Within a bundle, a {{{}-U{}}}/{{{}+U{}}} pair on
the same unique key is folded into a suppressed pair: the {{-U}} decrements the
count but keeps the record in state. The paired {{+U}} must then always count
again. Otherwise an outer row with two matches loses one, and deleting the
other match emits a null-padded row while a match still exists.
Example, with order 1 matching two shipments, {{S1}} and {{{}S2{}}}:
# One bundle: {{{}-U S1{}}}, {{+U S1'}} → the count must stay 2.
# {{-D S2}} → the count goes to 1. Only {{-D[order 1, S2]}} is expected, with
no {{{}+I[order 1, null]{}}}.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
