关于「百毫秒回馈:时延阈值胜过华丽特效」,《首分钟上手与粘性方程》给出的定量结论是什么?

结论:时延是分布,主战场在尾部。中位数只差 0.35 秒的两个版本,超过一秒的比例可以差十倍以上。骨架屏不缩短时延,却在示意模型里让离开比例下降 0.25;乐观更新只适合失败率低、可完整恢复的操作。串行请求让 P95 迅速越过一秒门槛,首个结果之前最多只容得下 2 次串行请求。特效可以有,但必须排在时延之后,而且可以关闭。

数据口径:首分钟漏斗是乘法:示意参数下,注册墙要走 9 步,每一万名访客只有 3,477 人看到第一个结果;先玩后存只要 2 步,有 8,924 人看到,约为前者的 2.57 倍,存档接受率达到 0.508 时两者账号数打平。 点按目标从 24 像素放大到 44 像素,示意误点率从 0.0865 降到 0.0017。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 0.25, 95

limits in seconds0.1, 1.0, 10.0; Doherty line 0.4
profile P50 P95 sigma <0.10.1-1 1-10 >10 <0.4
A local echo0.06 0.15 0.557 0.820 0.180 0.000 0.000 1.000
B fast API0.25 0.80 0.707 0.098 0.878 0.025 0.000 0.747
C slow API0.60 2.40 0.843 0.017 0.711 0.272 0.000 0.315
D cold start1.20 9.00 1.225 0.021 0.420 0.517 0.042 0.185
P95 budget check: share above 1.0 s must stay below 0.05
A local echo above 1 s 0.000 ok
B fast API above 1 s 0.025 ok
C slow API above 1 s 0.272 RED
D cold start above 1 s 0.559 RED
查看完整验算程序
# Chapter 2. p21: where a lognormal latency falls among the 0.1 s,
# 1 s and 10 s limits. A latency profile is set by its median P50 and
# its 95th percentile P95; sigma = ln(P95 / P50) / 1.6449. Shares come
# from the normal CDF (math.erf). The four profiles are illustrative.
import math

Z95 = 1.6449


def cdf(x, mu, sigma):
    return 0.5 * (1.0 + math.erf((math.log(x) - mu) / (sigma * 2 ** 0.5)))


def bands(p50, p95):
    mu, sigma = math.log(p50), math.log(p95 / p50) / Z95
    c1, c2, c3 = (cdf(x, mu, sigma) for x in (0.1, 1.0, 10.0))
    c4 = cdf(0.4, mu, sigma)
    return sigma, c1, c2 - c1, c3 - c2, 1.0 - c3, c4


if __name__ == "__main__":
    profiles = [("A local echo", 0.06, 0.15), ("B fast API", 0.25, 0.80),
                ("C slow API", 0.60, 2.40), ("D cold start", 1.20, 9.00)]
    print("limits in seconds: 0.1, 1.0, 10.0; Doherty line 0.4")
    print("profile        P50   P95  sigma  <0.1  0.1-1  1-10   >10"
          "  <0.4")
    for name, p50, p95 in profiles:
        s, a, b, c, d, e = bands(p50, p95)
        print("{:<13} {:>4.2f} {:>5.2f}  {:.3f} {:.3f}  {:.3f} {:.3f}"
              " {:.3f} {:.3f}".format(name, p50, p95, s, a, b, c, d, e))
    print("P95 budget check: share above 1.0 s must stay below 0.05")
    for name, p50, p95 in profiles:
        s, a, b, c, d, e = bands(p50, p95)
        print("  {:<13} above 1 s {:.3f}  {}".format(
            name, c + d, "ok" if c + d < 0.05 else "RED"))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《首分钟零摩擦上手、微交互成瘾回路与产品超高粘性留存工程母典》第 2 章「百毫秒回馈:时延阈值胜过华丽特效」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-379 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026379,
  author    = {Fan, Yafu},
  title     = {Zero-Friction First-Minute Onboarding, Micro-Interaction Habit Loops and High-Stickiness Interface Engineering},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论