Reading a record is not the same as completing its business effect. If a consumer commits before processing and then crashes, the restart can skip unfinished work. Processing before committing avoids that particular loss, but a crash between the effect and the commit can cause replay. The sink must tolerate that replay or participate in an appropriate atomic boundary. These tasks separate processed positions from committed offsets so the difference is visible. Process the contiguous records, then commit the processed position. The local process control stands in for successful business handling; it does not run an external payment or database. Check both the processed evidence and the resume position.
Inspect the topics, partition logs and group offsets before changing the cluster.