Latency tiers & cost — seconds, minutes, hours

Prompt (Feynman). A product manager says: "Make every mart second-fresh — storage is cheap." Explain in your own words the four cost buckets of a freshness tier (compute, storage, operations, people). Explain why the seconds tier can cost several times the hours tier even though the data is the same. Explain how you would model all three tiers, so that two engineers land within ±20% of each other. Use the revenue mart as the running example.

Rubric — your explanation should hit these (5–7)

• Freshness is the lever. The decision is "how fresh?", and the answer sets the bill. The bill is not linear.
• Four buckets. Compute, storage, operations and people. A tier changes all four at once.
• Compute dominates and grows steeply. Batch pays only while a job runs. Minutes and seconds tiers are effectively always-on clusters. Compute is the dominant cost at low latency.
• Storage grows mildly. Frequent commits mean more, smaller files plus compaction work. But the retained data volume is roughly the same. So storage rises less steeply than compute.
• Operations & people climb. A sub-minute job that must not lag is a 24/7 on-call commitment. Batch is a cron job.
• Same data, different bill. The seconds tier can cost several times the hours tier. The reasons are only always-on compute, small-file/compaction overhead and on-call. Justify it with numbers.
• Deterministic, reconcilable model. You have a workload and per-unit prices. Each component is a simple formula. Two independent models of the same workload agree within a tolerance (±20%) and keep the ordering seconds > minutes > hours.
• When not to. If hourly meets the SLA, seconds gives you nothing. Do not over-provision freshness.