Interview questions

The rate is the smallest factor. What counts as a unit, the billing increment, minimums and how well you use what you pay for all multiply it. A forecast of quantity × rate misses every one of them.

Read moreThe exchange rate between all the other unitsWhat pricing forbids

billed = ⌈q ÷ g⌉ × g × r: quantity rounded up to the billing increment, times the rate. A 5-second job billed in 60-second increments is charged for 12 times what it used.

Read moreThe exchange rate between all the other units

Every request costs the vendor the same front-end work, whatever its size. With a per-byte meter only, small requests would be almost free and could exhaust the service.

Read moreThe four things a vendor can meter

When p(read) × retrieval cost exceeds the storage saving, or when objects are small or short-lived and hit the tier's minimums, such as a 128 KB minimum size or a 90-day minimum duration.

Read moreTiering is a contract, not a mediumMinimums and floors

You pay for capacity-time, not work, so the effective price is the rate divided by utilization. At 20% utilization, B costs more per useful hour than a pricier engine running at 80%.

Read moreCompute: capacity-time, not work

Hundreds of MB. Request fees set an economic floor (about 100 KB just to break even on fees, larger once other costs count), and the 20–100 ms latency per request sets a physical floor; both point to the same region.

Read moreTwo floors that agree

Parallel reads can hide the latency of many small files, but not their request fees. Cost scales with files touched × queries, whatever the bytes.

Read moreTarget file size

Under pricing per byte scanned, the bill follows the data layout. Reading 1/40 of the columns, 1/30 of the partitions and a third of the remaining files thanks to clustering compounds to 3,600 times fewer bytes.

Read moreThe computable value of pruning

It depends on the billing clause. If you pay for the CPU, use lighter codecs for hot data. If you pay per compressed byte scanned, compress as hard as possible. If scanned bytes are counted uncompressed, compression only saves storage.

Read moreCompression as a lever

Physical data independence says the number of columns you read should not matter; pricing per byte scanned makes it the bill. Reading 200 columns when 5 are needed costs 40 times as much.

Read moreUnpruned scans and SELECT *

About the length of a cold start, tens of seconds. A timeout of 5–10 minutes is a purchase of lower latency, and it should be priced as one.

Read moreIdle warehouses

When the duty cycle, the fraction of time in use, is below 1 / k, where k is the premium for elastic capacity. Sparse traffic, a query every few minutes, is the worst case: always cold, never idle long enough to suspend.

Read moreScale-to-zero: a one-line break-even

The headroom needed is ramp rate × the loop's delay. Reactive scaling follows growth; spikes are absorbed by idle capacity, a queue, or shed load.

Read moreAutoscaling lag and over-provisioning

It needs all its nodes at once, so with a 5% chance per hour of any node being interrupted, it survives an hour with probability 0.95¹⁰⁰ ≈ 0.6%. Put the coordinator and state on on-demand machines, and checkpoint.

Read moreThe gang effect

At the percentile of forecast demand equal to the discount: the 40th for a 40% discount. An unused committed unit costs the committed price (0.6 of list), while a missing one costs only the discount (0.4), so the break-even is where demand falls below the commitment 40% of the time. And discount the headline saving for prices that fall over the term.

Read moreSizing from the shape of the lossA bet against price deflation

The cost signal arrives days or a month late, while cost can multiply within hours, and a slow control loop cannot regulate fast change. Add an inner loop of caps and estimates that act immediately.

Read moreNobody sees the outcome

Only under pricing per byte scanned: the bytes left after pruning are known from metadata. Under pricing by occupancy, cost depends on how the query behaves at run time.

Read moreUnder scan pricing, cost is a property of metadata

Query costs span about six orders of magnitude, so a cache with a 2.5% hit rate can save half the spend if it catches the expensive queries. Optimize dollars avoided, not hits.

Read moreHit rate is the wrong objective

Deleting risks an unbounded downside that is pinned on whoever deleted, against a small saving spread across everyone, so the default is to keep everything. Deleting safely also needs lineage showing what still uses the data.

Read moreLifecycle, tiering and retention

Each team optimizes its own line, for example by building private copies of shared tables, so data is duplicated and drifts.

Read moreHow chargeback fails

Those that come from structure: ratios, fixed latency floors, asymmetric losses. Not those that come from a particular number in the price list.

Read moreWhich verdicts survive a price change

Changing the workload. It is a negotiation with the people who generate the queries, not an engineering task.

Read moreThe escapes, ordered by price