The OpenLogReplicator ingestion method for Oracle uses a similar mechanism as XStream and PostgreSQL where we periodically need to confirm a checkpoint. This confirmation essentially marks a position in the event stream of where all committed transactions up to that point have been delivered successfully.

If a transaction starts (T1) and is followed by another (T2), the second transaction commits, and a checkpoint is sent. Confirming this checkpoint will cause all changes before the checkpoint related to T1 to be omitted from the event stream. This is a problem with OLR and not with Debezium.

The discussion with Adam seems to indicate that we may even be in a position where we may be able to resume mid-transaction.