retention

The most messages a partition keeps, at least one (DEFAULT_RETENTION unless set): once a write takes a partition past it, the partition drops its oldest messages, as a broker's size retention drops them, and the topic says so once, the first time. A read from an offset no longer kept starts at the oldest one kept, the offset gap a reader handles as it does on a broker (PartitionedSource.read). So the topics stay bounded however long the node runs: they cannot keep a message until its readers have read it, since a source knows no reader's position, which each keeps in its own checkpoint.