There is only one pass through Blackwater Ridge, so narrow that two mules walking abreast scrape the rock on either side. On this side lie six tea-growing villages; on the far side, the market. Nobody has ever measured how many mules the gap can swallow in a day — the mist sits low, the pass hides in cloud all year, and no one stands high enough to see the whole of it.

Old Tao was the youngest of the six caravan masters. When he took over his father’s mules, he did what every beginner does: he drove all thirty of them up at once. Eleven came home that night. The rest were jammed on the scree below the gap, cargo scattered, men spent.

The next year he changed his method. On the first trip he sent one mule. It came back at dusk with a wooden tally on its neck — the market clerk’s proof that it had crossed and delivered. So he sent two. Both tallies came home, so he sent four, then eight. He called it sounding the water: you don’t know how deep it is, so you feel your way down, one step at a time — but quickly.

On some trip, a mule failed to return. Old Tao neither cursed the mules nor sent men out to search. He did exactly one thing: he halved the size of the next trip. And from that number onward he added only one mule per trip. One. Even after ten safe trips in a row, still only one. People called him timid. He said: quick going up, slow coming down — the pass taught me that rule.

The strange part is that the other five caravans each arrived at the same rule on their own — add one, halve on loss — without ever discussing it. And over the years their shares of the pass ground themselves roughly even. Whoever got greedy lost mules sooner and was cut back harder; whoever stayed timid found his empty space filled, one mule at a time, by someone else. No magistrate enforced fairness. Fairness simply grew.

Then something odd happened. A local gentleman, offended by the mess on the scree, paid to build an enormous holding yard below the gap, big enough for a thousand loads. From that day, mules almost never went missing — there was always room to queue. But Old Tao noticed the tallies coming home later and later: first the same evening, then a day later, then three. No mules lost, yet the tea spoiled in the yard. The other caravans, still adding one mule for every trip without a loss, pushed harder and harder into the jam. Old Tao changed what he watched. Instead of only asking whether mules came back, he began recording how many days the tallies took. When that number crept well above its best, he cut back — even with not a single mule lost.

And one year the old soldier who kept watch above the gap began raising a red flag when the crowd grew thick. Caravans that saw the flag halved on their own, without waiting to lose anything. That year everyone’s tea arrived fresher than usual.

— By now you’ve probably recognized it: the unmeasured gap is the bottleneck bandwidth on a path; the number of mules Old Tao sends per trip is the cwnd (congestion window); the returning tallies are ACKs; one-two-four-eight is slow start; add-one-on-success, halve-on-loss is AIMD; the gentleman’s enormous yard is bufferbloat; Old Tao’s switch to timing the tallies is the delay-based family (Vegas, BBR); and the watchman’s red flag is ECN. This is TCP congestion control.

What it is, and why it matters

Congestion control answers a hard question: the sender cannot see how wide the narrowest link on the path is, nor how many others are competing for it right now. Nobody tells it. It can only probe with the packets it sends, inferring conditions from whether ACKs return and how fast. This is not flow control — flow control is the receiver explicitly saying “my buffer is full.” Congestion control faces a silent, shared, unmeasured bottleneck in the middle.

The classic mechanism has two phases. A new connection enters slow start, growing cwnd by one segment per ACK — effectively doubling every RTT — so it approaches the available bandwidth within a few round trips. On loss it shifts to congestion avoidance: from then on, one MSS per RTT, and an immediate halving on the next loss. That asymmetry between additive increase and multiplicative decrease is not arbitrary. It can be shown that AIMD-style rules are what let many independent, non-communicating flows converge simultaneously toward high utilization and a roughly fair split.

Bufferbloat is where this breaks in modern networks. Device buffers grew so large that congestion no longer drops packets — it just queues them. A loss-based algorithm senses nothing, keeps pushing, and latency climbs to hundreds of milliseconds or seconds: a link that “loses nothing but is agonizingly slow.” BBR and Vegas instead use RTT and measured delivery rate as signals; ECN lets routers mark congestion explicitly before dropping anything. Once you see this layer, you can explain why a long fat cross-region pipe won’t fill (cwnd is capped by RTT and loss together), why enlarging a buffer can make things worse, and why one bulk transfer can wreck the latency of interactive traffic sharing the same link.

Metaphor mapping

• The pass at Blackwater Ridge — the narrowest bottleneck link on the path, capacity unknown and shared
• Mules sent per trip — cwnd, the congestion window
• Wooden tallies brought home — ACKs
• One, two, four, eight — slow start, exponential probing
• A mule that doesn’t return — packet loss, read as a congestion signal
• Adding just one mule after a safe trip — congestion avoidance, additive increase per RTT
• Halving after a loss — multiplicative decrease
• Six caravans’ shares grinding themselves even — AIMD’s convergence to fairness without coordination
• The gentleman’s enormous holding yard — oversized device buffers, bufferbloat
• No mules lost, but the tea spoils — no loss, yet queueing delay explodes
• Timing how long tallies take to return — RTT as signal; Vegas / BBR-style delay-based control
• The watchman’s red flag — ECN, explicit congestion marking before drops
• Nobody ever measured the gap — the end-to-end principle: the network won’t tell you its capacity, you must infer it