故事 · 一个不挑系统、不挑分布的定律 Origin Story · A Law That Cares About Nothing Inside the Box
1961 年 MIT 的 John D. C. Little 严格证明了一件令人惊讶的事:只要系统稳定(进出守恒), 那么「平均排队人数 = 平均到达率 × 平均等待时间」就一定成立 —— 不管顾客怎么来、服务台几个、是先到先服务还是乱序, 统统不影响。它简单到像废话,却威力无穷:你不需要知道系统内部任何细节,光凭进出口的两个数,就能算出第三个。 精益生产把它翻译成车间语言:在制品越少,每件活穿过车间越快 —— 这就是看板、CONWIP 限 WIP 的理论根基。 In 1961, John D. C. Little at MIT gave a rigorous proof of something surprising: as long as the system is stable (what flows in eventually flows out), then average queue length = average arrival rate × average wait time — and that holds regardless of how customers arrive, how many servers you run, or whether the queue discipline is FIFO, LIFO, or random. It sounds almost trivial, yet the power is enormous: you do not need to know anything about the internals of the system — two numbers at the boundary give you the third. Lean translated this into shop-floor language: fewer items in process means each item flows faster — the theoretical bedrock of kanban and CONWIP.

1 流动系统:活进 → 排队 → 加工 → 离开 The Flow System: Arrive → Queue → Process → Leave

运行中Running

2 固定流速下:在制品越多,停留越久 At Fixed Throughput: More WIP Means Longer Wait

到达率 λ 固定时,W = L / λ 是一条直线:L 翻倍,W 也翻倍。所以限 WIP(压低 L)能直接、按比例地缩短前置时间 W。 At a fixed λ, W = L / λ is a straight line: double L, double W. Capping WIP (pushing L down) reduces lead time W proportionally and immediately.

3 现实里的利特尔定律 Little's Law in the Real World

缩短前置时间:车间在制品太多 → W 长、交期慢。砍掉一半 WIP,前置时间立刻减半 —— 不用提速也能更快交货。 Shorter manufacturing lead time: too much WIP on the floor → long W, slow delivery. Halve WIP and lead time halves on the spot — no speed-up required.
看板 / CONWIP:用卡片数量硬性封顶 WIP,靠的就是 L 一旦受限,W 必然下降的数学保证。 Kanban / CONWIP: a fixed number of cards puts a hard ceiling on WIP — and Little's Law guarantees that once L is capped, W must drop.
服务与排队:餐厅、客服、急诊室都用它估等待时间;店里 30 人、每分钟进 6 人 → 平均等 5 分钟。 Service queues: restaurants, call centers, and ERs all use it to estimate wait time. 30 customers inside, 6 arriving per minute → average wait of 5 minutes.
软件交付:把「需求」当工件,限制并行任务数(WIP)就能缩短从立项到上线的周期,敏捷看板正是此理。 Software delivery: treat each user story as a work item — capping the number of in-flight tasks (WIP) shrinks cycle time from kickoff to launch. That is exactly why agile kanban works.
一句话In One Line
L = λ × W 的威力在于它不挑系统:你管它内部多复杂,只要长期稳定,三个量就被这条等式锁死。 把它变形成 W = L / λ,结论一目了然:在流速 λ 不变时,前置时间 W 完全由在制品 L 决定。 于是精益不去逼工人「干快点」(提 λ 很难、还出错),而是限制在制品(压 L 很容易)—— 活堆得少,每件穿过系统的时间自然短。这就是为什么「少做一点」反而能「更快交付」,看板上的 WIP 上限正是这条定律的执行器。 The power of L = λ × W is that it does not care what is inside the box: as long as the system is stable in the long run, those three quantities are locked together. Rearrange to W = L / λ and the lesson is unmistakable: at a fixed throughput λ, lead time W is fully determined by WIP L. That is why lean does not push workers to "go faster" (raising λ is hard and tends to backfire); instead it caps the WIP (lowering L is cheap and immediate) — fewer items in the system means each item gets through faster. "Do less" really does mean "deliver sooner", and a kanban WIP limit is exactly the enforcement mechanism for this law.
常见误用Common Mistakes
用瞬时值套公式。L = λW 是长期平均关系,对刚启动、还没稳态的系统不成立。 Plugging in instantaneous values. L = λW is a long-run average — it does not apply to a system still warming up to steady state.
系统不稳定也照算。到达率持续大于处理能力时队列无限增长,定律失效;前提是进出守恒。 Applying it to an unstable system. If arrivals consistently outpace service, the queue grows without bound and the law breaks. Stability — inflow equals outflow in the long run — is the precondition.
以为提 λ 就能缩 W。提 λ 在 L 不变时反而不改善 W;真正缩 W 的是降 L(限 WIP)。 Assuming higher λ shortens W. Raising λ at fixed L does not reduce W. The lever that truly cuts W is reducing L — that is, capping WIP.

利特尔定律