5D Library › 零人工多站点容错架构、支付防欺诈状态机与被动现金流排队论 · English

第六章 流量分配、升级独立域名与现金储备

30 个站点之间分配付费流量

共享内核上的 30 个子站点,转化率各不相同,事先也不知道。每天有一笔固定的付费流量(这里设为 3,000 位访客)要分给它们。平均分配最简单,但会把大量流量浪费在早已看出转化很差的站点上。这是一个多臂老虎机问题。汤普森抽样是最早也最简洁的解法之一:对每个站点的转化率保持一个后验分布,每次按「从后验中抽一个值、谁最大给谁」的方式分配,于是越可能最好的站点得到越多流量,而尚不确定的站点仍有机会被试 (Thompson 1933; Russo et al. 2018)。

程序让 30 个站点的真实转化率从 Beta(2, 98) 抽取(平均 2%),连续 60 天,比较平均分配与按天批量的汤普森抽样。

推导

# Chapter 6. [p11] Thompson sampling of paid traffic across 30 sites.
# Thirty sites on the shared kernel convert visitors with unknown rates
# p_i drawn once from Beta(2, 98) (mean 2%). A daily budget of 3,000
# paid visitors is split across them for 60 days. UNIFORM gives each
# site 100 a day; THOMPSON (in daily batches) draws 200 samples from
# every site's Beta(1 + conversions, 1 + misses) posterior and gives
# each site the share of draws in which it came out best.
# 50 replications with fixed seeds. Reported: conversions per run,
# regret against always using the best site, and the share of traffic
# that went to the three best sites.
import random

SITES, DAILY, DAYS, REPS, DRAWS = 30, 3000, 60, 50, 200

def run(rng, rates, policy):
    conv = [0] * SITES
    miss = [0] * SITES
    sent = [0] * SITES
    for _ in range(DAYS):
        if policy == "uniform":
            alloc = [DAILY // SITES] * SITES
        else:
            wins = [0] * SITES
            for _ in range(DRAWS):
                d = [rng.betavariate(1 + conv[k], 1 + miss[k])
                     for k in range(SITES)]
                wins[d.index(max(d))] += 1
            alloc = [DAILY * w // DRAWS for w in wins]
        for i, n in enumerate(alloc):
            c = sum(rng.random() < rates[i] for _ in range(n))
            sent[i] += n
            conv[i] += c
            miss[i] += n - c
    top = sorted(range(SITES), key=lambda k: -rates[k])[:3]
    return sum(conv), sum(sent[k] for k in top) / sum(sent)

if __name__ == "__main__":
    out = {"uniform": [], "thompson": []}
    best = []
    for r in range(REPS):
        rng = random.Random(75011 + r)
        rates = [rng.betavariate(2, 98) for _ in range(SITES)]
        best.append(max(rates) * DAILY * DAYS)
        for pol in out:
            out[pol].append(run(rng, rates, pol))
    ideal = sum(best) / REPS
    print("   ideal (best site only): {:.1f} conversions".format(ideal))
    print("   policy     conversions   regret   share to top 3")
    for pol, res in out.items():
        c = sum(x for x, _ in res) / REPS
        s = sum(y for _, y in res) / REPS
        print("   {:9s} {:12.1f} {:8.1f} {:16.3f}".format(
            pol, c, ideal - c, s))
    print("\nReading: Thompson sampling moves most paid traffic to the")
    print("few sites that convert best and recovers most of the gap")
    print("to the ideal; uniform splitting wastes it on sites whose low")
    print("rate was obvious within days.")
#    ideal (best site only): 9905.2 conversions
#    policy     conversions   regret   share to top 3
#    uniform         3473.4   6431.9            0.100
#    thompson        9435.2    470.0            0.911
#
# Reading: Thompson sampling moves most paid traffic to the
# few sites that convert best and recovers most of the gap
# to the ideal; uniform splitting wastes it on sites whose low
# rate was obvious within days.

若一开始就知道哪个站点最好、把全部流量给它,60 天可得 9905.2 次转化。平均分配只得到 3473.4 次,与理想相差 6431.9;汤普森抽样得到 9435.2 次,只差 470.0,并把 0.911 的流量给了最好的三个站点。

什么时候把子站升级为独立域名

子站点表现好时,可以把它从共享内核的一个子路径升级为独立域名,以获得独立的品牌与搜索流量。设独立域名每年额外花费 6,000(域名、独立内容、外链与监控),能让每年 20,000 次自然访问提高 30%,每次转化价值 60。升级划算的条件是转化率高于 c*=6000/(0.30×20000×60)=1/60c^* = 6000/(0.30 \times 20000 \times 60) = 1/60。每个站点每天得到 100 位访客,程序比较三种升级规则:后验规则(P(p>c*)≥0.95P(p > c^*) \ge 0.95 时升级,≤0.05\le 0.05 时放弃)、计数规则(满 20 次转化就升级)、观测率规则(第 7 天起观测转化率一超过 c*c^* 就升级,60 天仍不够就放弃)。

推导

# Chapter 6. [p12] When to promote a site from a sub-path to its own domain.
# A standalone domain costs K = 6,000 a year (domain, separate content,
# links, monitoring) and is assumed to lift a site's 20,000 organic
# visits a year by u = 30%; a conversion is worth 60. Promotion pays
# when p > c* = K / (u 20,000 60) = 1/60. Each site receives 100
# visitors a day; its rate p ~ Beta(2, 98) as in p11. Rules, checked
# daily for up to 365 days:
#   BAYES  - promote when P(p > c* | data) >= 0.95 under a Beta(1, 1)
#            prior, drop when it is <= 0.05;
#   COUNT  - promote after 20 conversions (a common rule of thumb);
#   PEEK   - promote as soon as the observed rate exceeds c* after at
#            least 7 days, drop after 60 days below it.
# 2,000 simulated sites. Reported: wrong promotions (p < c*), missed
# good sites, mean days to a decision, undecided after a year.
import math
import random

K, U, VISITS, VALUE = 6000.0, 0.30, 20000, 60.0
C_STAR = K / (U * VISITS * VALUE)

def betacf(a, b, x):
    qab, qap, qam = a + b, a + 1, a - 1
    c, d = 1.0, 1 - qab * x / qap
    d = 1 / d
    h = d
    for m in range(1, 300):
        m2 = 2 * m
        aa = m * (b - m) * x / ((qam + m2) * (a + m2))
        d = 1 / (1 + aa * d)
        c = 1 + aa / c
        h *= d * c
        aa = -(a + m) * (qab + m) * x / ((a + m2) * (qap + m2))
        d = 1 / (1 + aa * d)
        c = 1 + aa / c
        h *= d * c
        if abs(d * c - 1) < 1e-12:
            break
    return h

def beta_cdf(x, a, b):
    lbt = (math.lgamma(a + b) - math.lgamma(a) - math.lgamma(b)
           + a * math.log(x) + b * math.log(1 - x))
    if x < (a + 1) / (a + b + 2):
        return math.exp(lbt) * betacf(a, b, x) / a
    return 1 - math.exp(lbt) * betacf(b, a, 1 - x) / b

def decide(rng, p, rule):
    conv = n = 0
    for day in range(1, 366):
        c = sum(rng.random() < p for _ in range(100))
        conv, n = conv + c, n + 100
        if rule == "BAYES":
            up = 1 - beta_cdf(C_STAR, 1 + conv, 1 + n - conv)
            if up >= 0.95:
                return True, day
            if up <= 0.05:
                return False, day
        elif rule == "COUNT":
            if conv >= 20:
                return True, day
        else:
            if day >= 7 and conv / n > C_STAR:
                return True, day
            if day >= 60:
                return False, day
    return None, 365

if __name__ == "__main__":
    print("   break-even rate c* = {:.4f}".format(C_STAR))
    print("   rule    wrong promotions  missed good  mean days  undecided")
    for rule in ("BAYES", "COUNT", "PEEK"):
        rng = random.Random(75012)
        wrong = missed = und = days = good = bad = 0
        for _ in range(2000):
            p = rng.betavariate(2, 98)
            d, t = decide(rng, p, rule)
            days += t
            if p > C_STAR:
                good += 1
                missed += d is not True
            else:
                bad += 1
                wrong += d is True
            und += d is None
        print("   {:6s} {:17.3f} {:12.3f} {:10.1f} {:10.3f}".format(
            rule, wrong / bad, missed / good, days / 2000, und / 2000))
    print("\nReading: the count rule promotes nearly every site, since")
    print("twenty conversions arrive within weeks at almost any rate;")
    print("peeking at the observed rate promotes almost a fifth of the")
    print("sites that do not pay. The posterior rule decides most sites")
    print("within weeks, with about one wrong promotion in seventeen;")
    print("sites near c* stay undecided longest, and for them the")
    print("choice hardly matters.")
#    break-even rate c* = 0.0167
#    rule    wrong promotions  missed good  mean days  undecided
#    BAYES              0.060        0.054       23.2      0.024
#    COUNT              0.995        0.000       20.3      0.003
#    PEEK               0.182        0.011       29.0      0.000
#
# Reading: the count rule promotes nearly every site, since
# twenty conversions arrive within weeks at almost any rate;
# peeking at the observed rate promotes almost a fifth of the
# sites that do not pay. The posterior rule decides most sites
# within weeks, with about one wrong promotion in seventeen;
# sites near c* stay undecided longest, and for them the
# choice hardly matters.

计数规则几乎把所有站点都升级了(不该升级的站点有 0.995 被升级),因为在任何转化率下,几周之内都会攒满 20 次转化。观测率规则把 0.182 的不该升级的站点升级了:早期的观测率噪声很大,一看到超过就升级,相当于反复偷看。后验规则平均 23.2 天作出决定,错误升级 0.060,漏掉该升级的站点 0.054,一年后仍未决定的只有 0.024——那些都是转化率接近 c*c^*、升不升级差别不大的站点。

「被动现金流」需要多少储备

现金流也是一个队列:订单的钱要过一段时间才到账,成本却每天都要付。程序用复合泊松过程描述它:每天平均 40 笔订单,单笔金额对数正态(平均 60、变异系数 4,少数高价牌照占大头),钱在下单 7 天后到账,3% 的订单在 30 天后退款或拒付,固定成本每天 1,800。期望的日净收入是 528.0。

推导

# Chapter 6. [p13] "Passive" cash flow as a compound Poisson process.
# Orders arrive at 40 a day (Poisson); an order's value is lognormal
# with mean 60 and CV 4 (a few licences dominate). Money reaches the
# account 7 days after the order (settlement delay); 3% of orders are
# refunded or charged back 30 days after the order. Fixed costs are
# 1,800 a day (servers, AI inference, part-time staff). Starting from a
# reserve R0, the program simulates 400 years of daily balances and
# reports the share of years in which the balance ever goes negative,
# with and without a revenue shock (orders halve for 90 days from day
# 30, as after a search-ranking loss).
import math
import random

LAM, MEAN, CV, COST = 40.0, 60.0, 4.0, 1800.0
SIG = math.sqrt(math.log(1 + CV * CV))
MU = math.log(MEAN) - SIG * SIG / 2

def poisson(rng, lam):
    k, p, L = 0, 1.0, math.exp(-lam)
    while True:
        p *= rng.random()
        if p < L:
            return k
        k += 1

def year(rng, r0, shock):
    inflow = [0.0] * 400
    out = [0.0] * 400
    bal, low = r0, r0
    for d in range(365):
        lam = LAM / 2 if shock and 30 <= d < 120 else LAM
        for _ in range(poisson(rng, lam)):
            v = rng.lognormvariate(MU, SIG)
            inflow[d + 7] += v
            if rng.random() < 0.03:
                out[min(399, d + 30)] += v
        bal += inflow[d] - out[d] - COST
        low = min(low, bal)
    return low < 0, bal

if __name__ == "__main__":
    rng = random.Random(75013)
    print("   expected net per day: {:.1f}".format(
        LAM * MEAN * 0.97 - COST))
    print("   costs before the first settlement: {:,.0f}".format(7 * COST))
    print("   reserve R0   P(negative) plain   P(negative) with shock")
    for r0 in (15000, 20000, 30000, 60000, 90000):
        a = sum(year(rng, r0, False)[0] for _ in range(400)) / 400
        b = sum(year(rng, r0, True)[0] for _ in range(400)) / 400
        print("   {:10,d} {:19.3f} {:24.3f}".format(r0, a, b))
    print("\nReading: a business that earns on average still needs a")
    print("reserve, because cash arrives a week late and the costs do")
    print("not wait. A few weeks of costs covers ordinary noise; a")
    print("three-month traffic shock needs several times as much. The")
    print("queue of money is as real as the queue of requests.")
#    expected net per day: 528.0
#    costs before the first settlement: 12,600
#    reserve R0   P(negative) plain   P(negative) with shock
#        15,000               0.028                    0.995
#        20,000               0.000                    0.980
#        30,000               0.000                    0.970
#        60,000               0.000                    0.318
#        90,000               0.000                    0.000
#
# Reading: a business that earns on average still needs a
# reserve, because cash arrives a week late and the costs do
# not wait. A few weeks of costs covers ordinary noise; a
# three-month traffic shock needs several times as much. The
# queue of money is as real as the queue of requests.
储备 60,000 时的 8 条年度余额路径:无冲击(蓝)与第 30—120 天订单减半(紫)。

平均每天赚 528.0 的生意,在第一笔回款到账之前就要先付 12,600 的成本。储备 15,000 时,平常年份有 0.028 的概率出现负余额;储备 20,000 以上,平常年份就不再出现负余额。但遇到订单减半三个月的冲击(例如搜索排名下跌),储备 30,000 时 0.970 的年份会出现负余额,60,000 时仍有 0.318,90,000 时才降到 0。「被动现金流」的「被动」,是以储备为前提的。复合泊松过程的破产概率是保险数学的经典问题,这里只用模拟给出具体参数下的数字。

假说 · 尚未验证

作者的诠释(尚未验证): 作者的站群收入在单个搜索引擎上的集中度很高,三个月的流量冲击并非罕见;因此储备应当按冲击情景而不是平常年份设定,至少相当于 50 天的固定成本。什么会证明这一判断错:若过去三年的真实流量记录中,没有任何连续 30 天的订单下降超过三成,那么冲击情景可能过于悲观,储备可以按平常年份设定。

小结

汤普森抽样把付费流量集中到最好的几个站点,与理想只差 470.0 次转化,而平均分配差 6431.9 次。升级独立域名应当用后验规则:计数规则几乎全部升级,观测率规则错误升级 0.182。现金流需要储备:平常年份 20,000 已够,三个月的订单减半需要 90,000。

参考文献

Russo, Daniel J., Benjamin Van Roy, Abbas Kazerouni, Ian Osband, and Zheng Wen. 2018. “A Tutorial on Thompson Sampling.” Foundations and Trends in Machine Learning 11 (1): 1–96.
Thompson, William R. 1933. “On the Likelihood That One Unknown Probability Exceeds Another in View of the Evidence of Two Samples.” Biometrika 25 (3/4): 285–94.

Fan, Y. (2026). Zero-Touch Multi-Site Fault Tolerance, Payment Anti-Fraud State Machines and Passive Cashflow Queuing: Receipt-Hash Deduplication, Self-Referral Blocking, Host-Header Routing and Availability Bounds [零人工多站点容错架构、支付防欺诈状态机与被动现金流排队论] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090201 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).