In the current version, we deduplicate Records in the Writer's Queue. However, this makes it necessary to hold a log point for each read Record (ever). In large collection this regresses performance significantly (e.g. by bloating the backlog size to the number of records read, i.e. for our Telegram collection north of 150M records).

However, given a stateful reader, that keeps track of the stream position, we could circumvent the above described situation.