As observed during the demo, the loop used by the replayer to fetch from kafka items to replay isn't reliable, in the sense that it commits immediately after fetch, but can fail before having replayed them. We should delay the commit to after the replay has been committed to the target storage.