Lifecycle, tiering and retention

Why is deletion the least-used cost lever, and when should derived data be kept?

Retention is structurally different from every other lever here. Storage cost is an integral over time, so deletion is the only mechanism that reduces the integral and not just the rate. Everything else makes the accrual slower. Deletion stops it.

That makes it surprising that it is the least used lever. The reason is not technical. The cost of deleting something that turns out to be needed is unbounded, and it falls on an identifiable person. The cost of keeping it is a few dollars a month, spread over a budget that nobody reads.

Cheaper storage does not change this. It makes it stronger, because it shrinks the visible side of the asymmetry and leaves the fear intact.

The mechanical precondition is informational too: you cannot delete safely without knowing who reads the asset. So retention policy depends on usage telemetry and lineage. Attribution is not only a reporting convenience here. It is the enabling mechanism for the only lever that permanently reduces cost.

For derived data the decision has a clean form: store it if storage over the horizon is cheaper than the recompute cost times the number of times you expect to need it. This is the materialization break-even from materialization, read backwards.

Retention held for reprocessing is one of several priced correctness decisions, with replication factor, snapshot depth and quality-check compute: money buying a safer position on the cost-versus-safety trade-off, and rarely visible on a cost dashboard.