5D Library › 企业瓶颈诊断、产品线自蚕食审计与自动化改造回报边界 · English

第二章 推入瓶颈的工作、沟通通道与波动

「二次方膨胀」的说法要改正

常听到的说法是:在非瓶颈盲目上自动化或加人,在制品与沟通摩擦会「呈二次方膨胀」。核对之后要分开两件事。第一,非瓶颈本身的提速不会增加在制品(第一章中反而减少了);在制品上升,是因为上游更快之后把更多工作推入了瓶颈。第二,被推入的瓶颈的在制品不是二次方增长,而是双曲线增长:单服务台队列的平均在制品是 L=ρ/(1−ρ)L = \rho/(1-\rho),负荷 ρ\rho 接近 1 时趋于无穷。真正按二次方增长的是团队内部的沟通通道数 n(n−1)/2n(n-1)/2 (Brooks 1975)。

推导

# Chapter 2. p03: What pushing more work into the bottleneck does.
# (a) M/M/1 work in progress L = rho / (1 - rho) is hyperbolic in the
#     load, not quadratic: it diverges as rho -> 1.
# (b) Faster upstream teams release more leads into conversion (3
#     closers, 10 leads/day each). A lead waiting w days still closes with
#     probability p0 * exp(-w / T): leads go cold. For M/M/c the queue
#     wait has E[exp(-s W)] = (1 - Pw) + Pw * (c mu - lam) / (c mu - lam + s).
# (c) Coordination channels in a team of n people: n (n - 1) / 2.
from math import factorial

def erlang_c(lam, c, mu):
    a, rho = lam / mu, lam / (c * mu)
    s = sum(a ** k / factorial(k) for k in range(c))
    top = a ** c / (factorial(c) * (1 - rho))
    return top / (s + top)

C, MU, P0, T = 3, 10.0, 0.5, 1.0

def push(lam):
    pw = erlang_c(lam, C, MU)
    gap = C * MU - lam
    wq = pw / gap                                  # mean wait, days
    lq = lam * wq
    lt = (1 - pw) + pw * gap / (gap + 1 / T)       # E[exp(-W/T)]
    return lq + lam / MU, 24 * wq, lam * P0 * lt

if __name__ == "__main__":
    print("   (a) rho    L=rho/(1-rho)   0.5-matched quadratic 4*rho^2")
    for r in (0.5, 0.8, 0.9, 0.95, 0.98):
        print("       {:.2f}  {:13.2f}   {:10.2f}".format(
            r, r / (1 - r), 4 * r * r))
    print("   (b) leads/day  WIP  wait (h)  closed sales/day")
    best = max((push(l)[2], l) for l in [x / 2 for x in range(40, 60)])
    for lam in (20.0, 24.0, 26.0, 28.0, 29.0, 29.5):
        L, w, sold = push(lam)
        print("       {:9.1f}  {:5.1f}  {:7.1f}  {:8.2f}".format(
            lam, L, w, sold))
    print("       most sales at {:.1f} leads/day: {:.2f} per day".format(
        best[1], best[0]))
    print("   (c) team size  channels")
    for n in (4, 6, 8, 12):
        print("       {:9d}  {:8d}".format(n, n * (n - 1) // 2))
#    (a) rho    L=rho/(1-rho)   0.5-matched quadratic 4*rho^2
#        0.50           1.00         1.00
#        0.80           4.00         2.56
#        0.90           9.00         3.24
#        0.95          19.00         3.61
#        0.98          49.00         3.84
#    (b) leads/day  WIP  wait (h)  closed sales/day
#             20.0    2.9      1.1      9.60
#             24.0    5.0      2.6     10.89
#             26.0    7.5      4.6     11.03
#             28.0   15.1     10.5      9.91
#             29.0   30.1     22.5      7.70
#             29.5   60.1     46.5      5.22
#        most sales at 25.5 leads/day: 11.06 per day
#    (c) team size  channels
#                4         6
#                6        15
#                8        28
#               12        66

把双曲线与一条在 ρ=0.5\rho = 0.5 处对齐的二次曲线 4ρ24\rho^2 对比:负荷 0.80 时在制品是 4.00 件,二次曲线只有 2.56;负荷 0.90 时是 9.00 件对 3.24;负荷 0.95 时是 19.00 件对 3.61;负荷 0.98 时是 49.00 件对 3.84。离饱和越近,双曲线越比二次方危险。

推得越多,成交越少

转化团队有三名顾问,每人每天处理 10 个线索。上游提速后每天释放进来的线索越来越多。线索在队列中等得越久越冷:等待 ww 天之后成交的概率是 p0e−w/Tp_0 e^{-w/T}。每天放入 20.0 个线索时,队列在制品 2.9 个,平均等待 1.1 小时,每天成交 9.60 单;放入 26.0 个时成交 11.03 单;放入 28.0 个时,在制品 15.1 个、等待 10.5 小时,成交反而降到 9.91 单;放入 29.5 个时,在制品 60.1 个,等待 46.5 小时,成交只剩 5.22 单。成交最多的放入速度是每天 25.5 个(11.06 单),远低于团队的名义产能。

转化团队的每日成交与在制品随上游放入线索速度的变化(p03)。

与此同时,非瓶颈团队加人还有另一种代价:8 人团队有 28 条两两沟通的通道,加到 12 人就是 66 条。人数增加一半,通道增加一倍多。

注

放行速度跟着瓶颈走。 约束理论称之为「鼓、缓冲、绳」:上游放行的节奏由瓶颈决定,而不是由上游自己的产能决定 (Goldratt and Cox 1984)。在本例中,这意味着投流的预算应按转化团队每天能消化的线索数来定,而不是按投流团队能产出的线索数来定。

非瓶颈可以帮忙:降低波动

非瓶颈的改善也有正当的价值,但通常不是来自提速,而是来自让流向瓶颈的工作更平稳。金曼公式给出排队等待与波动的关系 (Kingman 1961),连接方程给出一站离开流的波动如何由它的到达波动与服务波动混合而成 (Hopp and Spearman 2008)。

推导

# Chapter 2. p04: A non-bottleneck CAN help - by smoothing, not speed.
# Acquisition (one server) feeds conversion (3 closers, the bottleneck).
# Kingman/Sakasegawa queue wait for G/G/c:
#   Wq = (ca2 + cs2)/2 * rho^(sqrt(2(c+1)) - 1) / (c (1 - rho)) * te
# and the linking equation for the variability of departures
#   cd2 = 1 + (1 - rho^2)(ca2 - 1) + rho^2 (cs2 - 1) / sqrt(c).
# Leads arrive at 25 per day with ca2 = 1.5 (bursty campaigns).
from math import sqrt

LAM, CA2 = 25.0, 1.5

def station(lam, ca2, c, mu, cs2):
    rho = lam / (c * mu)
    te = 1 / mu
    wq = (ca2 + cs2) / 2 * rho ** (sqrt(2 * (c + 1)) - 1) / (
        c * (1 - rho)) * te
    cd2 = 1 + (1 - rho ** 2) * (ca2 - 1) + rho ** 2 * (cs2 - 1) / sqrt(c)
    return wq, cd2, rho

def line(mu_acq, cs2_acq, cs2_conv=1.0):
    w1, cd2, r1 = station(LAM, CA2, 1, mu_acq, cs2_acq)
    w2, _, r2 = station(LAM, cd2, 3, 10.0, cs2_conv)
    return 24 * w1, cd2, 24 * w2

if __name__ == "__main__":
    cases = (("base", 40.0, 2.0),
             ("acquisition twice as fast", 80.0, 2.0),
             ("acquisition half as variable", 40.0, 0.5))
    b = line(40.0, 2.0)
    print("   case                          acq wait  cd2   conv wait (h)")
    for name, mu, cs2 in cases:
        w1, cd2, w2 = line(mu, cs2)
        print("   {:<29} {:7.2f}  {:5.3f}  {:7.2f}  ({:+.0%})".format(
            name, w1, cd2, w2, w2 / b[2] - 1))
#    case                          acq wait  cd2   conv wait (h)
#    base                             1.75  1.695     4.63  (+0%)
#    acquisition twice as fast        0.24  1.549     4.38  (-5%)
#    acquisition half as variable     1.00  1.109     3.63  (-22%)

投流团队按批次上线广告,交出的线索时多时少。原方案中,转化团队的平均等待是 4.63 小时。把投流团队的速度提高一倍,它自己的等待从 1.75 小时降到 0.24 小时,但转化团队的等待只降到 4.38 小时(-5%)。不提速,只把投流交接的波动减半(例如固定时段交接、每批大小相同),离开流的波动系数平方从 1.695 降到 1.109,转化团队的等待降到 3.63 小时(-22%)。

假说 · 尚未验证

作者的假说(尚未验证): 在内容获客型的小团队中,把投流到转化的交接改成固定时段、固定批量,比把投流团队的产能提高一倍更能缩短线索的首次响应时间。什么会证明它错: 在按周随机切换的实验中,前者的首次响应时间不短于后者。

小结

非瓶颈提速本身不增加在制品;在制品膨胀是因为更多工作被推入瓶颈,而它按 ρ/(1−ρ)\rho/(1-\rho) 双曲线增长,比二次方更危险。按二次方增长的是沟通通道。推入过多会让线索变冷,成交反而下降。非瓶颈对瓶颈最大的帮助,往往是让交给它的工作更平稳。

参考文献

Brooks, Frederick P. 1975. The Mythical Man-Month: Essays on Software Engineering. Addison-Wesley.
Goldratt, Eliyahu M., and Jeff Cox. 1984. The Goal: A Process of Ongoing Improvement. North River Press.
Hopp, Wallace J., and Mark L. Spearman. 2008. Factory Physics. 3rd ed. McGraw-Hill/Irwin.
Kingman, J. F. C. 1961. “The Single Server Queue in Heavy Traffic.” Mathematical Proceedings of the Cambridge Philosophical Society 57 (4): 902–4.

Fan, Y. (2026). Enterprise Bottleneck Diagnostics, SKU Cannibalization Audits and Automation Return Bounds: Theory-of-Constraints Queueing Networks, Margin Dilution Matrices and Value-Based Advisory Pricing [企业瓶颈诊断、产品线自蚕食审计与自动化改造回报边界] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090217 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).