skip to content

Your reporting fleet never drops below forty machines of one profile and triples for three days each month - how much of it would you commit to?

level: seniorimportance: should knowfreq 45%

answer

  1. draw the fleet over a month
  2. find the line it never crosses
  3. the average is the trap
  4. commit the floor, meter the spike
  5. ladder terms against roadmap confidence

basics

~20 s

Commit to the floor and meter the spike. The forty machines that are up every hour justify a committed rate; the eighty that exist three days a month run about a tenth of the hours and would cost more committed than metered.

solid answer

~50 s

The floor is the only part worth committing to, and even then not all of it at one term length. Using illustrative units - metered 1.0 per machine-hour, committed 0.6 charged every hour - a 30-day month costs 34,560 entirely metered, 23,040 with the floor committed and the spike metered, and 51,840 with the whole peak fleet committed. Committing the spike is worse than committing nothing, because those eighty machines exist for 72 of 720 hours, about 10%, against a break-even of 60%. I would then split the floor by confidence rather than by size: the part that survives whatever is on the roadmap takes the long term, and the rest takes a shorter one or stays metered. I would also want the floor measured across a full quarter, since a floor observed over two weeks is a guess.

code

pseudocode · 19 lines
pseudocode
// illustrative units, not any provider's price list
meteredRatePerHour   = 1.0
committedRatePerHour = 0.6      // billed for every hour of the term
hoursInMonth = 720

floorMachines = 40              // up every hour
peakMachines  = 80              // extra, present 3 days a month
peakHours     = 72              // 10% of the month

allMetered   = floorMachines * hoursInMonth * meteredRatePerHour
             + peakMachines  * peakHours    * meteredRatePerHour      // 34560

floorCommitted = floorMachines * hoursInMonth * committedRatePerHour
               + peakMachines  * peakHours    * meteredRatePerHour    // 23040

allCommitted = (floorMachines + peakMachines) * hoursInMonth * committedRatePerHour  // 51840

cheapest = min(allMetered, floorCommitted, allCommitted)              // 23040
// note: allCommitted exceeds allMetered - committing above the floor adds cost

go deeper

for a junior

Know that a commitment is sized against the capacity that is always running, and that capacity which appears for a few days a month is normally paid for by the hour instead.

for a middle

Do the comparison on monthly totals for at least three mixes, and be able to show that committing above the floor can cost more than committing nothing.

for a senior

Demand a full business cycle of measurement, separate floor from spike instead of reasoning about the average, and size the committed band against the roadmap rather than the chart alone.

for a principal

Own the laddering policy: how much of the fleet's bill may be non-cancellable, across which terms, who reviews it when an architecture changes, and what the uncommitted band is insuring against.

## Find the floor before you price anything A commitment is sized against the line the fleet **never drops below**, not against its average and not against its peak. So the first step is a chart of running machines over at least one full business cycle - for a month-end reporting service, at least a month, and preferably a quarter so that an unusual close does not set the line. Three numbers come out of that chart: - the **floor**: 40 machines, present every hour; - the **spike**: an extra 80 machines, present for three days a month, which is 72 hours out of roughly 720; - the **average**, which is about 48 machines and is the most dangerous number on the chart, because committing to the average means committing to 8 machines that only exist one day in ten. ## The arithmetic on three mixes Using invented units - a metered rate of 1.0 per machine-hour and a committed rate of 0.6 per machine-hour charged for every hour of the term - over a 720-hour month: | Mix | Committed part | Metered part | Monthly total | |---|---|---|---| | Everything metered | - | 40 x 720 + 80 x 72 | 34,560 | | Floor committed, spike metered | 40 x 720 x 0.6 = 17,280 | 80 x 72 = 5,760 | **23,040** | | Whole peak fleet committed | 120 x 720 x 0.6 = 51,840 | - | 51,840 | The middle row is a third cheaper than the first. The third row is worse than buying nothing at all, which is the point worth internalising: **a commitment sized above the floor does not reduce a discount, it adds cost.** The spike machines run about 10% of the hours against a break-even of 60%, so each one bought on commitment is paid for around the clock and used one day in ten. ## Why not commit to the whole floor either The floor is the right *shape* to commit to, but the right *quantity* is a judgment about confidence, not about the chart: 1. **Roadmap risk.** If a re-platforming is planned, some of that floor is going to disappear inside the term. A commitment does not disappear with it. 2. **Measurement risk.** A floor measured over two weeks can be a seasonal artefact. Commit to the part you have watched for a quarter. 3. **Instrument risk.** If the commitment is tied to a specific machine shape, a change of profile strands it; if it is expressed as spend per hour that any matching usage draws down, it survives more change. Providers differ, and this materially changes how much of the floor is safe to cover. The practical form is a **ladder**: cover the part of the floor you are confident about on a longer term, cover a further slice on a shorter term, and leave a deliberately uncommitted band on metered capacity. The uncommitted band is not waste - it is the part of the bill that can shrink when the architecture does, and it is what lets you renew into a changed shape instead of renegotiating out of a wrong one. ## What to do with the spike The spike stays metered by default. If some of the month-end work is regeneration or backfill that can be stopped and restarted, that portion is a candidate for reclaimable capacity - bought cheap precisely because the provider can take it back - but that is a different bargain with a different risk, and the interactive part of the close window should not be on it. One boundary worth naming out loud in an interview: a discount instrument is a **billing** construct. Sizing it correctly tells you what you will be charged; it does not tell you that 80 extra machines of that profile will be obtainable on the day of the close. That is a separate question with its own answer. ## How to present the answer An interviewer is listening for four moves: you asked for the shape over a full cycle; you separated floor from spike rather than reasoning about the average; you did the utilisation comparison rather than comparing rates; and you sized the committed band against how far ahead you can honestly see. Getting to a specific number matters far less than showing that the number came from the chart and from the roadmap rather than from the discount table.

  • Why is committing to the fleet's average size worse than committing to its floor?
    Because the machines between the floor and the average are not there most of the time. Each one above the floor is billed at the committed rate around the clock while existing for a small share of the hours, so it is bought above its break-even utilisation. The average is a statistic about the chart; the floor is the part of the chart that is always true.
  • How would you decide the size of the deliberately uncommitted band?
    By how much of the floor could plausibly disappear inside the term - the fleet a planned re-platforming removes, a customer that could churn, a workload that might move tiers. Cover what survives every one of those, ladder a second slice on a shorter term, and leave the rest metered. The band is priced insurance against your own roadmap.

saying these in an interview costs you the question

  • Committing to the fleet's average size rather than its floor
  • Committing the spike because the discount percentage looks large
  • Signing the longest available term for every machine in the floor
  • Assuming an unused commitment simply stops appearing on the bill
  • Treating a recurring three-day spike as steady demand