What elasticity forbids

Which elasticity mistakes are forbidden, and who holds the risk in each escape?

What this article forbids

Elasticity without a duty cycle. Every mechanism above resolves against d and the premium, and usually neither is measured.
Spot without a restart granularity. The discount means nothing next to an unbounded loss term.
Commitment at the mean of a forecast, given an asymmetric loss function.
The belief that autoscaling replaces headroom. It replaces sustained over-provisioning, and nothing on the time scale of a spike.

The escapes and their prices

Who holds the risk	You pay
You hold it	in idle
The vendor holds it	the on-demand or per-invocation premium
You take the vendor's leftover risk	in interruption handling plus correlated unavailability
You sell your flexibility forward	in forecast error plus deflation exposure

There is no fifth option. The pattern is familiar from systems design: trade-offs are paid in three currencies (waiting, wasted work, weakened guarantees). Money is admitted as a fourth. Cost buys position on every other axis, so wherever a list of currencies exists, money extends it by one.

Case, resolvedConstructed case · Tomás at Brightloom

A spot job that never finishes

Priya proposed moving a 6-hour, 100-node job to spot, had bought a three-year commitment sized to the peak week, and set the development warehouse to scale to zero although analysts queried every few minutes. Where it started: Elasticity →

The spot job. 100 nodes, 5% interruption per hour, a job that needs all of them: survival per hour is 0.95¹⁰⁰ ≈ 0.6%, and the expected time to the first interruption is about 12 minutes. He moves the coordinator and the state to on-demand machines, runs the workers on spot, and checkpoints every 5 minutes (average loss 2.5 minutes per interruption). The job finishes, at a cost near 40% of the on-demand price (illustration).
The commitment. It was sized to the peak week. Priya will commit next year near the P40 of the forecast (the quantile equal to her 40% discount), and will prefer one-year terms, since the three-year discount erodes as list prices fall.
The development warehouse. Analysts are active every four or five minutes, so it is always cold and never idle. He replaces it with a small always-on warehouse. With a duty cycle near 100% and a premium of 3, going elastic could never win: d × k = 3.