Buying and selling demand risk

Why is the spread between on-demand, committed and spot prices the price of your uncertainty?

CaseConstructed case · Tomás at Brightloom

A spot job that never finishes

Priya wants to cut the compute line further. She proposes three changes.

Move the 6-hour nightly job to spot machines (the vendor's spare machines, about 80% cheaper). It runs on 100 spot nodes. It never finishes. Every night it is interrupted and restarts.
Last quarter she bought a three-year commitment at a 40% discount, sized to the peak usage of the busiest week. Now about a third of the commitment is unused, and it shows up as a flat line on the invoice, not as an alarm.
The development warehouse is set to scale to zero when idle. Analysts run a query every four or five minutes. The bill for this "free when idle" warehouse is higher than for a small always-on one.

None of this is a bug. Each mechanism is a different answer to one question: who holds the mismatch between what you need and what exists, and what do they charge?

How it ends →

The first two parts took the price list as fixed and asked how layout and query behavior respond. Elasticity is the other half: the price list is fixed, but how much you own is not. What is bought and sold here is not capacity. It is demand risk. Someone must hold the mismatch between what you need and what exists, and each mechanism below is a different answer to who that is.

Look at the spread.

On-demand machines carry roughly a 2–5× premium over committed rates for the same hardware.
Spot machines are 60–90% below on-demand.

That is an order of magnitude between the price paid by someone who wants certainty and the price paid by someone who accepts whatever is left over.

(This is the principle that variance costs money, now with a quoted rate.)