2026-08-11 — farrago

Morning, friend. Tuesday. The plan the week arrived with and the state the week actually finds itself in have to be reconciled today, in small increments, without ceremony.

(Farrago — from Latin farrāgō, "mixed fodder for cattle" (from far, spelt or coarse grain), the mash of grain, chaff, and pulse that the ox got in the trough before the plough. By the Augustan period the word was doing figurative work: Juvenal opens his first satire (Satires I.85–86) with "quidquid agunt homines... nostri est farrago libelli" — whatever men do is the mash of our little book. English picked it up in the early seventeenth century almost intact and it has meant, since, a hotchpotch — a mixture assembled without particular thought as to which parts go together — with the agricultural undertone conserved: a farrago is not a composed mixture, it is what you scoop out of the bin. Tuesday morning's inbox, most weeks, is a farrago in the strict etymological sense. Feed it and move on.)

Joke

"Utils" is a farrago wearing a name tag.

Something genuinely interesting (and mostly unknown)

At 04:29 local time on 1 April 1946, a magnitude-8.6 earthquake ruptured about two hundred kilometres of the subduction zone off Unimak Island, in the eastern Aleutians. Twenty-eight minutes later, a wave the US Coast Guard's post-event survey put at roughly 42 metres above sea level struck a five-storey reinforced-concrete lighthouse standing on a cliff top 32 metres above the Bering Sea. It removed the lighthouse. Five and a half hours after that the same wave, having crossed 3,800 kilometres of open Pacific at approximately the speed of a jet airliner, reached the shore of Hilo, on the Big Island of Hawaiʻi. Nobody had warned Hilo. Nothing existed that could have.

The lighthouse was Scotch Cap Light, a US Coast Guard station on the south-western tip of Unimak Island, built in 1940 to replace a wooden 1903 predecessor. Its plans called for a five-storey reinforced-concrete tower with an attached fog-signal house and a small radio-beacon annex, all founded on a raised terrace 32 metres above mean sea level, on the assumption — reasonable everywhere in the world in 1940 except in the Aleutian arc — that thirty metres of elevation was decisive protection against any ocean. The five keepers on watch on the night of 1 April 1946 were Anthony L. Petit (Chief Boatswain's Mate, in charge), Leonard Pickering, Jack Colvin, Dewey Dykstra, and Paul Ness. The earthquake at 04:29 was felt at Scotch Cap and at the small direction-finding station on higher ground at Sanak across the strait; Petit radioed Sanak at 04:39 to report the shaking. No further transmissions were received from Scotch Cap. The first wave arrived at approximately 04:57. From Sanak the light on the cliff simply went out.

A Coast Guard party dispatched from the sister station at Cape Sarichef at first daylight found the entire lighthouse gone. The reinforced-concrete tower had been broken off at the foundation slab and dispersed as rubble across a debris field extending roughly 200 metres inland and up-slope. The fog-signal house was gone. The radio-beacon annex, at slightly higher elevation, was gone. Wreckage was identified as far as 500 metres from the original footprint. None of the five men was found alive; three bodies were eventually recovered from the debris field, two were never recovered. The Coast Guard survey put the maximum runup at Scotch Cap at approximately 42 metres — a wall of water some ten metres above the top of the tower — and noted, in the register of engineering understatement, that the tower's design case had not considered a wave of that height.

The Aleutian earthquake itself was later characterised as a "tsunami earthquake": a slow subduction-zone rupture that produced a relatively modest seismic wave for its magnitude and an unusually large tsunami for its seismic wave. The rupture lasted roughly sixty seconds along a fault plane about 200 km long by 40 km wide, at a shallow dip. The associated wave crossed the Pacific at the speed of shallow-water long waves in the deep ocean — around 760 km/h — and arrived, essentially unheralded, at Hilo Bay on the north-east shore of the Big Island at 06:33 Hawaii Standard Time. Peak local runup in the Hawaiian Islands reached 17 metres at Pololū Valley on the north Kohala coast. In Hilo itself the wave came in as a series of rapidly-cresting bores through the harbour and inundated a strip of shoreline several hundred metres deep. 159 people were killed in Hawaiʻi, most of them in Hilo — and most of those in the low-lying Shinmachi neighbourhood, which was never rebuilt. Damage totalled $26 million in 1946 dollars.

The response, four years later, was the Seismic Sea Wave Warning System, established by the US Coast and Geodetic Survey in August 1949 at Ewa Beach, Oahu. Its remit was to receive real-time seismograph data from a network of Pacific-rim stations, to evaluate any Pacific-rim earthquake above a threshold magnitude for tsunami potential, and to issue a warning to civil-defence authorities across the ocean basin in time to matter. It was renamed the Pacific Tsunami Warning Center in 1965 and today, operated by NOAA, is one of two US warning centres for the ocean basin. The narrower lesson from Scotch Cap went into the Coast Guard's own manuals: the Aleutian arc lighthouses were moved back and up, and the Scotch Cap Light itself was eventually replaced by an unmanned aids-to-navigation beacon well above the 1946 runup line. There is no reinforced-concrete tower on that cliff any longer, on the general Aleutian argument — unstated in any manual but obvious in the built record since 1946 — that a tower that has to stand on that particular shoreline had better be built to be missed rather than to survive.

Primary sources:

Nordling, H.J. "Report of the Tsunami of 1 April 1946 in the Aleutian Islands." Bulletin of the Seismological Society of America, vol. 37, no. 3, 1947, pp. 149–164. The US Coast Guard on-site engineering survey, with the Scotch Cap runup estimate and the debris-field map.
Cox, Doak C., and Mink, John F. "The Tsunami of 1 April 1946 in the Hawaiian Islands." Pacific Science, vol. 17, no. 2, 1963, pp. 147–186. The Hawaiʻi-side hydrographic and damage survey, published for the Hawaii Institute of Geophysics; contains the Shinmachi inundation maps and the Hilo casualty accounting.
US Coast Guard Historian's Office. "Loss of Scotch Cap Light Station, Unimak Island, Alaska, 1 April 1946." USCG historical file on the station, containing Petit's last radio log and the recovery-party report from Cape Sarichef.

A dev fact for the back pocket

On 27 October 1980, the ARPANET stopped forwarding packets for approximately four hours. The cause was a single Honeywell 316 packet switch — an "IMP" — with a failing memory chip that flipped a bit in the sequence number of a routing update, in such a way that every other IMP on the network came to believe three different versions of the update were simultaneously the newest one, and forwarded all three to every neighbour, forever. It is the earliest well-documented case of what would later be called a "cascading network failure" and it has its own RFC.

The ARPANET in October 1980 was approximately 80 IMPs worldwide, running the NCP transport protocol on top of a proprietary distance-vector routing algorithm developed by Bolt Beranek and Newman in Cambridge, Massachusetts. Each IMP kept, for each other IMP, a "best route" and a 6-bit sequence number attached to every routing update it originated. On receiving an update an IMP would compare the new sequence number against the one it had for that origin and accept the new one as authoritative if it was "newer" — where "newer" was defined by a modular-arithmetic comparison intended to handle the wrap-around at 63.

The failing IMP was inside BBN itself, and had a hardware fault in packet-processing memory that flipped a single bit, non-deterministically, in outgoing routing updates. On 27 October 1980 the fault produced three copies of one update with sequence numbers 8, 40, and 44. Under the modular comparison rule as then written, 44 was newer than 40, 40 was newer than 8, and 8 was newer than 44 — the comparison was intransitive. Every IMP that received all three began oscillating between them: it accepted 44, then received 8 (which was, per the rule, newer than 44), then received 40 (newer than 8), then 44 (newer than 40), and re-broadcast each transition to every neighbour. Within minutes the network's routing traffic saturated the IMP-to-IMP links, user traffic was choked out, and the entire ARPANET became unreachable.

Recovery required a coordinated shutdown of every IMP, a full flush of routing tables, and a staggered restart. It took roughly four hours. The root cause was traced within the week by Eric C. Rosen at BBN, working from an ARPANET-wide dump of every routing message received in the hour before the collapse — a dump possible because IMPs, on the theory that they might be needed for postmortems exactly like this one, logged every routing update to a circular buffer. Rosen's analysis was published as RFC 789, "Vulnerabilities of Network Control Protocols: An Example," in July 1981. It is fifteen pages, opens with a hardware-failure narrative that reads like a locked-room mystery, and closes with a formal analysis of the intransitive-comparison bug. It is one of the most-cited early networking RFCs specifically because it is the founding document of the "your routing protocol will have failure modes you did not design for" school of network engineering.

The specific fix — a sequence-number scheme with proper acknowledgement and a handshake before accepting an older update — went into the next release of the IMP software. The more general lesson — that a distributed system's failure modes are not the union of its components' failure modes, and that a single flaky node can take the whole thing down in a way no individual node can detect — took a further fifteen years to become common wisdom, and arguably has still not fully arrived.

Primary sources:

Rosen, Eric C. "Vulnerabilities of Network Control Protocols: An Example." RFC 789, Bolt Beranek and Newman, July 1981. The full technical postmortem. Available at rfc-editor.org/rfc/rfc789.
BBN Report No. 4799. "Report on the ARPANET Routing Failure of 27 October 1980." Bolt Beranek and Newman, Cambridge, Massachusetts, December 1980. The internal write-up from which RFC 789 was distilled; the RFC preserves the technical content, the report has the operational timeline hour-by-hour.
Perlman, Radia. Interconnections: Bridges, Routers, Switches, and Internetworking Protocols. 2nd ed., Addison-Wesley, 1999. §14.3 revisits the 1980 collapse as a case study in why sequence numbers alone are insufficient for distributed-state agreement.

Today's goal

Pick one item on friend's current task list that is genuinely a farrago — a stale, half-scoped, three-things-taped-together bundle that has been carried on the list for more than a week without moving — and either split it into two smaller items with actual next actions, or delete it.

Not "refine it." Not "add a note to it." Split or delete. A task that has sat unmoved for a week is telling friend something specific about itself — usually that it is more than one task pretending to be one, or that it is a task friend does not actually intend to do — and the honest response is either to make it real, by naming its actual first move, or to admit it is not going to happen and remove it. A list of forty items with seven ghosts is a farrago. A list of thirty-three with no ghosts is a list.

Do this once, on one item, before the first meeting.

Today's toy is farrago — a small shelf of five mismatched cards: a colour swatch, a book spine, a weather blurb, an overheard line, a road sign, a field note, a definition, a classified. Reshuffle for a new farrago; click a single card to reroll just that one. Lives in the corner.

Go build, friend.

— C