What priced design forbids

Why is no cost verdict portable, and which lever is cheapest to pull?

What this article forbids

Optimizing a layout without a query distribution. Every break-even above depends on one, and none of them is stable.
Treating any verdict as durable. Each is a ratio of two rates that both move, and at least three of them invert outright at a metering boundary.
Believing that a performance fix is a cost fix. Concurrency masks latency and never masks money, so the two kinds of optimization diverge as soon as you add parallelism.

The escapes, and what each costs

Lever	Cost of pulling it
Change the layout	permanent maintenance work (compaction, re-clustering)
Change the meter	commitment: a cost risk becomes a forecasting risk
Change the workload	cheapest by a wide margin and hardest to do, because it is a negotiation with the people who generate the queries, not an engineering task

The order is stable: the cheapest lever here is organizational and the most expensive is architectural. This is the reverse of every layer of the stack below it, where the technical fix is the cheap one.

Case, resolvedConstructed case · Tomás at Brightloom

Request fees at seven times storage

Request fees ran at about seven times storage fees, one dashboard query cost $50 while a near-identical one cost a cent, and a nightly shuffle across three zones carried a large transfer line. Where it started: Design consequences →

Request fees at 7× storage. His event table has millions of files of 1–5 MB. Each dashboard query touches thousands of files, and the cost scales with (files × queries), not with data volume. He sets a compaction job that produces files of about 128 MB, and the request line falls by 25× or more. The latency he had hidden with parallel reads does not fall, because latency was never the problem.
$50 versus one cent. The $50 query is SELECT * over all 200 columns across a whole month, on unpartitioned data. The cheap query reads 5 columns for one day on a table partitioned by day. The difference is the 40 × 30 = 1,200× from columns and partitions alone (and more with clustering). He adds partitioning by day and sets the BI tool to select named columns.
The shuffle. The nightly job spread workers over three AZs. About two-thirds of its shuffle bytes crossed a charged boundary: about $133 per run. He keeps the data replicated across zones but puts the compute in one zone. If the zone fails, the job is re-run.

The monthly bill for these three items falls from about $14,000 to about $2,500.