Metric Provenance

A capacity plan arrives citing "620 million viewers". A reviewer asks whether the platform must serve 620 million people at once. It must not: that is season-long reach, a count of distinct humans over seven weeks. The number that sizes infrastructure is peak concurrency, two orders of magnitude smaller.

Nobody lied. Three numbers describing one platform - reach, cumulative views, peak concurrency - came from three instruments built for three audiences, and only one is a planning input. Once a number is in a slide it travels without its instrument.

Why it matters

An architect's inputs are almost all second-hand, and the cost of not checking them is asymmetric. Substituting a larger-definition number for a smaller one over-provisions and looks like prudence, so it is never caught; substituting a vendor maximum for an achieved rate under-provisions and is caught during the event it was meant to survive. Provenance is also the cheapest authority available in a review: "that is a peak rate from a 2021 blog post and our planning number has to be a sustained rate from this quarter" ends a disagreement in one sentence, without requiring seniority.

Implementation patterns

• Three fields per number: instrument, date, denominator. "25.3 million concurrent viewers, platform-reported peak, 10 July 2019, one live event." A number that cannot be written this way does not enter the document.
• Classify before comparing. Reach counts people over a period; cumulative views counts play events; concurrency counts simultaneity. Two numbers in different classes are never a ratio.
• Prefer the filed over the published. A filing is written under liability and a product page to sell; where both exist and differ, the filing is the planning number and the gap is information.
• Name peak and sustained as separate fields - one column called traffic will mix them inside a quarter.
• One test, always: can this number exceed the population that produced it? Reach and views can; concurrency, utilisation and share cannot, so a concurrency figure above plausible audience size is a mislabelled view count.

Industry example

Akamai's own figures illustrate the discipline. Its FY2023 annual filing describes more than 4,100 edge points of presence in approximately 130 countries and nearly 750 cities; the FY2025 filing gives more than 4,300 points of presence and roughly 700 cities; an earlier company blog post said more than 1,000 cities. All are accurate as published and the counting rule moved underneath them, so quoting "1,000+ cities" today quotes a definition, not a fact.

On throughput, Akamai announced in March 2021 that delivered traffic peaked at 200 Tbps, having passed 100 Tbps in October 2019; a later post described well over 100 Tbps around the clock. Those do not conflict - one is a peak rate and one a sustained load.

Failure scenarios

• Reach used as concurrency. Capacity sized 100x too high, approved because nobody objects to a margin, and the unit economics quietly fail.
• A vendor maximum used as an achieved rate. A 400 Gb/s link delivering 40% of that under the real message pattern, found in the first production run.
• A peak used as a sustained rate in a cost model, so the bill lands 3x over forecast and the overrun is blamed on growth.

Trade-offs

Insisting on provenance occasionally removes a figure everyone liked, leaving a hole where confidence used to be. That hole is the correct state of knowledge, and a reviewer who has not internalised that reads the rigorous document as the weaker one.

When not to use it

Skip it for numbers that do not bear load: an illustrative figure in a narrative, a rough analogy for scale, a number deciding which of two options gets an afternoon. The test is whether a 3x error would change a decision. Reserve the three fields for the two or three numbers a design rests on - a team that demands provenance for everything gets it for nothing.

Interview question

Q: "Here is a capacity brief citing a platform's published viewership figures. Which would you refuse to use, and what would you do if the only available number were the wrong class?"

What a strong answer covers: separating reach and views from peak concurrency and saying why only the last sizes infrastructure. Then, for the hard part, deriving bounds rather than guessing - an upper bound from audience population, a lower bound from the last measured event - and presenting a design valid across that band with the measurement that collapses it.

Quick check

Quiz: A brief cites 620 million viewers and 25 million concurrent for one platform. Which sizes the delivery tier? Peak concurrency - the worst simultaneous second; reach counts distinct people over a period, so the two are different instruments and their ratio means nothing.

Flashcard: Which three fields accompany every load-bearing number, and what test catches the commonest mislabelling? - Instrument, date and denominator; ask whether the number could exceed the population that produced it.