A reporting service has a flat weekday base and a sharp month-end peak - which purchase posture suits each part of that demand shape?
answer
- the shape has layers
- price each layer separately
- commit only to the floor
- the peak runs too few hours
- cheap for two opposite reasons
basics
~20 sBuy the always-on base with a term commitment, serve the month-end peak with metered on-demand capacity, and put only interruption-tolerant extra work on reclaimable capacity. Price each layer of the shape separately rather than picking one posture for the fleet.
solid answer
~60 sI would draw the fleet size over a whole month first, because the shape decides the answer, not the workload's name. The part that is up every hour of every day is the **floor**, and that is what a term commitment is for: a lower rate in exchange for being billed for the committed level whether I use it or not. The month-end peak exists for a few days a month, so committing to it would mean paying a discounted rate around the clock for capacity that runs a tenth of the hours - it stays on metered on-demand capacity even though it is perfectly predictable. **Reclaimable capacity** - rented cheap because the provider keeps the right to take it back at short notice - is only for work that can be stopped and restarted, such as backfill or report regeneration, never the interactive tier. The two discounts are cheap for opposite reasons: on a commitment I took the risk, on reclaimable capacity the provider kept the right to take it away.
go deeper
Know that the same machine hour can be bought three ways: pay as you go with no promise, a lower rate bought with a promise to keep paying for a term, or cheap capacity the provider may take back.
Be able to decompose a fleet into a floor, a predictable peak and deferrable work, and say which posture each layer takes and why utilisation - not predictability - decides whether committing wins.
Show that you would measure the floor over a full business cycle before signing, and that you check what the covered workload's roadmap does to that floor inside the term.
Frame it as risk allocation across the fleet: how much of the bill you are willing to make non-cancellable, how that band is laddered across terms, and who reviews it when the architecture moves.
## What you are actually choosing between A compute hour is sold three ways. Every large provider offers all three, under its own names, and the names matter far less than which side of the deal carries which risk. - **Metered on-demand capacity.** You start a machine, you are billed for the time it runs, and billing stops when you stop it. Highest unit price, no commitment, no interruption. The provider carries the risk that the machine sits idle. - **A term commitment.** You promise a level of usage, or of spend, for a fixed term - usually measured in years, sometimes a single year - and receive a lower rate for it. The committed level is billed for the whole term whether you consume it or not. **You** carry the risk. - **Reclaimable capacity.** You rent capacity the provider currently has spare, at a deep discount, and it keeps the right to take the machine back at short notice. The provider carries no idle risk at all, which is exactly why the price is low; you carry the interruption. So the two discounted postures are cheap for **opposite** reasons, and that is the sentence worth having ready in an interview: a term commitment is discounted because *you* accepted the risk of not using it, and reclaimable capacity is discounted because the *provider* kept the right to take it back. Mixing those up is the classic error, because it leads to treating reclaimable capacity as a cheap always-on tier. ## Decomposing the demand shape Plot machines running against time for at least one full business cycle - for a reporting service that means a month, because the month-end close is part of the shape. Three layers fall out: 1. **The floor** - the line the fleet never drops below, including nights and weekends. This is capacity that is up close to all of the hours. 2. **The predictable peak** - extra capacity that appears on a known schedule and then goes away. Predictable in *timing*, but present for a small share of the hours. 3. **Deferrable work** - jobs that have to finish eventually but not at a particular minute, and that can be restarted without anybody noticing. Each layer has a posture that fits it: | Layer of the shape | Posture | Why it fits | What it costs when the shape changes | |---|---|---|---| | Floor, up nearly all hours | Term commitment | High utilisation is what converts a lower rate into a lower bill | The committed level keeps billing if the floor drops | | Predictable peak, up a few days a month | Metered on-demand | You pay only for the hours the capacity exists | Nothing - the bill follows the shape | | Deferrable, interruption-tolerant work | Reclaimable capacity | Deepest discount, and the work can absorb being stopped | The work takes longer, or waits, when capacity is taken back | ## Why the predictable peak still goes on metered capacity Predictability is not the test - **running hours are**. A term commitment bills for every hour of the term, so it only wins for capacity that is up most of the time. If the committed rate is some fraction of the metered rate, the break-even sits at exactly that fraction of the hours. Machines that exist for three days in a month are up about a tenth of the hours; for a commitment to pay there, the committed rate would have to be about a tenth of the metered rate, and term discounts are nothing like that deep. Commit to the peak and you pay a discounted rate around the clock for capacity you use one day in ten. ## Where reclaimable capacity actually belongs Reclaimable capacity is a purchase posture, not a tier you can silently substitute. It fits work whose progress can be thrown away and redone, or whose deadline is hours rather than seconds. Putting the interactive part of a reporting service on it means accepting that some of the fleet disappears at a moment the provider chooses, which is a workload-design problem in its own right and a separate subject from the purchase decision. ## Three checks before you sign anything 1. **How long has the floor been the floor?** A floor measured over two weeks is a guess; over two quarters it is a fact. 2. **What is on the roadmap for the covered workload?** A commitment outlives most architecture decisions, and a re-platforming that removes the covered fleet does not remove the bill. 3. **What is the instrument expressed in?** Some commitments are tied to a specific machine shape and some to an amount of spend per hour that any matching usage draws down. Providers differ here, and the flexible form survives a change of shape far better. One thing the choice does **not** settle: a discount instrument is a billing construct. Whether a machine of the shape you want is actually available when you ask for it is a different question with a different answer.
- Why are a term commitment and reclaimable capacity cheap for opposite reasons?On a term commitment the discount is payment for risk **you** took: you are billed for the committed level for the whole term whether you use it or not. On reclaimable capacity the discount is payment for a right the **provider** kept: it can take the machine back at short notice. One fails when your demand shrinks, the other when the provider's spare capacity does.
- The month-end peak is perfectly predictable. Why is that not a reason to commit to it?Because a commitment is billed for every hour of the term, not for the hours you use. Predictability tells you *when* the capacity is needed; only its share of total hours tells you whether a committed rate is cheaper. Capacity that exists three days a month runs about a tenth of the hours, far below any realistic break-even.
saying these in an interview costs you the question
- Committing to the month-end peak because the peak is predictable
- Treating reclaimable capacity as simply a cheaper always-on tier
- Buying a single posture for the whole fleet to keep billing simple
- Assuming a longer term is automatically the better buy
- Believing a commitment discount also guarantees the machine will be there