Read the author list again. That is essentially the entire founding pantheon of the modern Internet signing one document — because in 1998, the End-to-End Research Group genuinely feared the Internet was about to collapse under its own success. RFC 2309 is their joint plea to router vendors: stop building dumb tail-drop FIFOs.

The problem it confronts is subtle. TCP only learns about congestion by losing packets. A traditional router queue fills up, then drops every arriving packet until it drains — "tail drop." This sounds fair, but it produces two pathologies the RFC names explicitly:

• Lock-out: a few aggressive flows monopolize the queue while well-behaved flows starve.
• Full queues: tail-drop only signals congestion when the buffer is already saturated, so steady-state queue occupancy hovers near 100%, adding latency to every packet for no throughput benefit. This is the phenomenon Jim Gettys would later popularize as bufferbloat — but RFC 2309 diagnosed the mechanism a decade earlier.

The RFC's core recommendation is Active Queue Management (AQM): routers should drop (or mark) packets before the queue is full, probabilistically, as a deliberate congestion signal. The reference algorithm it endorses is Sally Floyd and Van Jacobson's RED (Random Early Detection): once average queue depth crosses a low threshold, drop incoming packets with a probability that ramps up linearly to the high threshold. Random selection means the drop probability for a flow is proportional to its share of the bandwidth — implicit fairness, no per-flow state required.

Two design decisions deserve attention:

• The RFC carefully distinguishes congestion control (what TCP does, end-to-end) from congestion avoidance (what routers should help with). It refuses to let the network do TCP's job, but insists the network must give TCP timely signals.
• It introduces the concept of "unresponsive flows" — UDP traffic that ignores drops — and warns that without router-level policing, these will eventually drive the Internet into congestion collapse. This single paragraph motivated two decades of subsequent work on fair queueing and ECN.

The history is bittersweet. RED was deployed widely but tuned terribly; its parameters (min_th, max_th, w_q, max_p) are notoriously sensitive to link speed and traffic mix, and most operators set them once and forgot them — or disabled AQM entirely. Bufferbloat got worse through the 2000s as cheap RAM made consumer routers ship with multi-second buffers and no AQM at all.

The redemption came around 2012 with CoDel (Controlled Delay, Nichols & Jacobson) and FQ-CoDel, which replaced RED's queue-length heuristic with a sojourn-time measurement that needs no tuning. RFC 7567 formally updated 2309 in 2015 to recommend these newer algorithms. If you have ever run tc qdisc on Linux and seen fq_codel as the default, or noticed your home router from the last five years finally has decent latency under load — that is RFC 2309's recommendations finally being honored, twenty-some years late.

Read it today for the framing more than the algorithm. The taxonomy of queue pathologies, the responsive-vs-unresponsive distinction, and the insistence that signaling latency matters more than buffer size are still the cleanest available statement of why your network feels slow.