The Multiplicative Law of Fan-Out Tail Latency and Hedged Requests: How Percentile Latency Amplifies When One Page Makes Many Calls: chapter answers
Website owners commonly believe that if every backend service keeps its slow requests to one per cent, the page experience cannot be too bad. This book overturns that belief with probability models. A page waits for all its calls, so its latency is their maximum, and one slow request in a hundred at each service becomes a slow-page probability of 0.634 with a hundred parallel calls. The number comes from "The Tail at Scale" by Dean and Barroso in Communications of the ACM, 2013; after checking the original, the book corrects the blueprint: it is a consequence computed under independence for a…
DOI pending · monograph page goes live after publication · 5DL-B-326 · 中文版
- Chapter 1 · The Multiplicative Law: How One Slow Call in a Hundred Becomes Sixty-Three Slow Pages in a Hundred
Chapter 1 recomputes the multiplicative law: one slow request in ten thousand across two thousand calls still gives slow pages with probability 0.181, and a page 99th percentile allows each call a slow share of only 0.0… - Chapter 2 · The Page Is a Maximum: Order Statistics and the Last Few Leaves
Chapter 2 writes the page as a maximum and tests independence against the measured table: the independent model puts the ratio of "ninety-five per cent of leaves returned" to "all returned" at 0.087, while the measureme… - Chapter 3 · When Independence Fails: Correlated Slowness, Bad Machines and Synchronised Background Work
Chapter 3 presents both directions of correlation: bad epochs lower the slow-page probability to 0.4033, fixed bad machines visited without replacement raise it to 0.6740, and synchronised background work presses it dow… - Chapter 4 · Hedged Requests: Sending a Second Copy at the 95th Percentile
Chapter 4 covers hedged requests: sending a second copy at the single call's 95th percentile adds 0.0500 extra requests and cuts the page miss probability to 0.0100; when slowness lies in the request itself, hedging fai… - Chapter 5 · Queueing Tails: Utilisation Stretches the Tail, and Hedges Are Load Too
Consider utilisation first. At a utilisation of 0.30, a call's mean response time is 7.1 milliseconds and its 99th percentile 32.9 milliseconds; at 0.70 they rise to 16.7 and 76.8 milliseconds; at 0.90 to 50.0 and 230.3… - Chapter 6 · Budget Allocation: Splitting a Page Percentile Across Every Call
Chapter 6 splits the page budget across calls, and splitting it unequally by cost saves 0.0597 of the capacity cost. - Chapter 7 · The Classics East and West: Half the Journey at Ninety, and the Law of the Minimum
The result is an almost literal confirmation. In the light-tailed model with a hundred leaves, the ninetieth leaf returns on average at 8.78 milliseconds and the last at 17.56 milliseconds, a ratio of exactly 0.500: at… - Chapter 8 · A Fan-Out Audit of a Paid Web Service, and a Plan for Refutation
Chapter 8 gives an offline fan-out audit for the pages of a paid web service, the distilled Fan-Out Tail Latency Calculator and a pre-registered test. Marked boxes separate published evidence, runnable Python calculatio…
Illustrative model calculation, not a live quote