Retention Cost
The storage bill for keeping a log replayable, which is set by retention multiplied by throughput multiplied by the replication factor.
Streaming cost has three components and retention is the one that grows quietly. Compute is always-on and visible in the bill from day one. Cross-zone network transfer is significant and frequently unmodelled. Retention is throughput × duration × replication, and each of those is set once and rarely revisited.
Thirty days of a topic carrying five hundred gigabytes a day at a replication factor of three is forty five terabytes of hot storage for one topic — an amount that would receive scrutiny as a database and receives none as a retention setting.
The tension is that retention is also the reprocessing window, so cutting it to save money silently removes the ability to replay. That trade should be made deliberately, per topic, against the recovery scenario each supports.
Two structural remedies. Compaction for changelog topics bounds size by key cardinality rather than by time, which is often a large reduction. Tiered storage, where older segments move to object storage while remaining readable, decouples retention from broker cost and is the single most effective change for an estate keeping long histories.