
The notification service reads the topic once at startup and exits, and every restart re-sends everything. Build a consumer loop that subscribes under a group with a defined starting position, polls with a timeout, exposes full record metadata, and stops cleanly when woken.
A backfill job needs one partition from a specific offset, and all anyone has is subscribe and hope. Read chosen partitions without joining a consumer group, seek to offsets, beginnings, ends and timestamps, report positions and end offsets, and refuse mixing subscribe with assign.
One poll returns thousands of records and a single malformed value kills the consumer. Cap every poll, deserialize values into typed records, report a malformed record with its partition and offset and move past it, survive exceeding the poll interval, and make sparse polls wait for data.
Support needs a replay tool: replay a topic from an offset, a time or the beginning, over chosen partitions and for a set count per partition, printing full metadata. Report and skip malformed records, stop at end-of-partition with the shortfall, and never touch a live consumer group.