Metadata and control-plane operations

Why does every structure that cuts data cost raise metadata cost?

Operations on the control plane (listing, creating, committing, snapshotting) scale with structural cardinality: object count, partition count, snapshot count, task count. They do not scale with data volume.

This gives the most useful general shape here. Nearly every mechanism that reduces data cost by adding structure increases metadata cost. The two costs sit on separate meters that move in opposite directions as one parameter changes. Partition granularity is the standard case:

• more, smaller partitions → better pruning → less data scanned (data meter goes down);
• more, smaller partitions → more files, more listing, more metadata (metadata meter goes up).

So the best value is in the interior, between the extremes, and it can be computed. "More pruning is always better" is false.

Where the control plane is rate-limited and not priced, the price is still paid, in latency, retries and complexity (per-prefix limits). A cost is never removed. It moves to another place.