Window glass gets stiffer faster the cooler it gets, and its own glow fades out at about 525 °C, inside the band where it can still relieve its internal stresses. Glassblowers read that slope with their bodies, and the craft hands the dark stage to a kiln on a schedule. Two habits for running software follow.
The bubble on the end of the blowpipe was meant to become a vase. In a beginners' glassblowing course in New York in the winter of 2004, the ethnographer Erin O'Connor's partner had heated it, sat down at the bench and asked O'Connor to blow. The partner had overheated the glass without meaning to, and O'Connor, not noticing, blew "with the force of industrial bellows." The egg-shaped bubble "bulged into a large amorphous blob."
Her partner took the blob to the glory hole, the barrel-shaped furnace where glass is reheated while it is being worked, rested the pipe on its yoke and turned it while the glass warmed. Then she looked away to chat. When she looked back, the glass was flopping at the end of the pipe. She stopped turning and tried to lift it out, which was the wrong response: the turning is what keeps the glass on center. By the time the piece was free of the furnace's doors it had stretched into a tube nearly two feet long.
Two experienced glassblowers who had been working at another glory hole were cleaning up their station. "You can save it, you can save it!! Make a cookie for it!!!" one called out as they walked by. One dropped a fresh gather of glass onto the marver and pressed it into a disc with a paddle, and the other lowered the long piece onto it, giving the failed vase a bottom. O'Connor, by her own account, had no idea what they were talking about. The sentence she later wrote about that day is the subject of this essay: "The novice, though she can search, cannot see the problem, i.e. that to be solved, without corporeally already knowing the answer."
What the two passers-by could see, and the beginners could not, was a curve.
Glass does not melt at one temperature the way ice does. Edgar Dutra Zanotto, of the materials engineering department at the Federal University of São Carlos in Brazil, puts the melting range of window glasses at 1,400 to 1,500 °C. Below that, the glass simply grows stiffer as it cools, continuously, with no step anywhere. In a 1998 paper in the American Journal of Physics he wrote that stiffening down for a typical modern window glass, mostly silica with soda and lime, using the standard empirical fit for glass viscosity, the Vogel-Fulcher-Tammann equation, with constants estimated from the glass's composition:
log η = −2.6 + 4077.7 / (T − 254.7)
Here η is the viscosity in pascal-seconds and T is the temperature in degrees Celsius.
The glass trade marks the hot end of that curve with named points. A 2015 lecture on annealing, part of a glass-processing course that Lehigh University publishes online, gives their definitions: the working point is where the viscosity is 1,000 pascal-seconds, and the softening point is where it reaches 10 to the power 6.65, about 4,500 times stiffer. On Zanotto's curve the working point falls at 983 °C and the softening point at about 696 °C. The stretch between them is about 290 degrees wide.
What makes the craft hard is the slope of the curve, and the slope changes as the glass cools. Near 1,100 °C, about 50 degrees of cooling doubles the viscosity of this glass. At the working point it takes about 37 degrees. Near the softening point it takes about 14. The same 20 degrees of cooling that raises the viscosity by about a third at 1,100 °C nearly doubles it at 800 °C. As the glass cools the window closes, and it closes faster the further it has closed.
That is the physics inside O'Connor's failed vase. Overheated glass sits further up the temperature scale, runnier than the person blowing expects, so the same breath does more. And a glassblower who decides what to do with a gather and then reaches for a tool is shaping glass that has already moved down the curve from the glass they looked at. The beginners in New York had no feel for how fast that was happening. The two passers-by did, and they carried it in their bodies.
The same curve explains a story Zanotto called "a quite universal concept", and shows why it is wrong. He wrote his paper because, over two decades of teaching, students and colleagues kept telling him that "800-year-old stained-glass windows of 12th century cathedrals were thicker in their lower part", so glass must flow at room temperature. "It is interesting to note," he wrote, "that no one knew the source of this information." He first took it for a Brazilian myth, then heard it from a colleague in Argentina, and a referee told him it was widespread in the United States. Even the Encyclopaedia Britannica alluded to glass deforming permanently at ordinary temperatures.
His curve cannot be run down to room temperature, because it predicts an infinite viscosity at 254.7 °C. So he worked from a better-behaved glass, germanium dioxide, whose viscosity rises more slowly than window glass's as it cools. Its relaxation time at room temperature came out at 10 to the power 32 years, and window glass would take longer still: "well beyond the age of the Universe." A year later Zanotto and Prabhat Gupta revised the estimate for window glass down by several orders of magnitude, using better data, and reached the same conclusion: window glass cannot flow at room temperature in human time scales. Run the other way, his numbers say a typical medieval glass would have to be held at about 414 °C to flow noticeably in 800 years. Museums agree: glass vases from thousands of years ago "remain undeformed."
So why would some old panes be thicker at the bottom, if they are? Zanotto's answer is the glassblower: "ancient window glasses were blown into cylinders that were split and flattened manually. Hence, the pieces were not uniform in thickness and some lower parts could be thicker than the upper parts." A fact about a craft's hands had been read for decades as a fact about physics. The thick edge, where there is one, is a record of a person working against the same curve.
If the curve is the physics, the craft is the set of moves for staying on the right part of it, and most of them buy time. The glory hole puts heat back in and pushes the glass back up the curve. Turning keeps it centered while it is soft. And because the window is short and closing, the work is shared: "glassblowing is always a collaborative endeavor," O'Connor writes.
What her article describes is how a person comes to read that curve without thinking about it. The beginner, she writes, "practices, reflects upon her practice, adjusts her corporeal dispositions to better align with instructors' expectations, and, so transformed, once again practices, which allows for incorporation and thus the recession of this new corporeal disposition into unconsciousness." In the moment, "what 'solves' the problem at hand is not a moment of cognitive reflection, but rather the response of the spontaneous dispositions that are equipped to manage the situation." The two glassblowers who rescued the vase responded on the spot. Where the beginners had come to "a stalemate of action," they saw "a problem to be solved," and each of them offered a different solution.
This is a different account of expertise from the one engineers tend to carry around, in which the expert simply knows more rules. Much of the glassblower's knowledge is a model of rates: how fast this glass, at this glow, is moving toward the point where it stops forgiving. O'Connor's novice cannot see the problem because seeing it takes that model, and no list of facts about glass supplies it.
Hot glass reports its own temperature in light. In 1847 John William Draper, professor of chemistry in the University of New York, set out to pin down where that light begins, because the authorities disagreed. Sir Isaac Newton, he noted, had fixed the temperature at which bodies become self-luminous at 635 °F, Sir Humphry Davy at 812°, Wedgwood at 947° and Daniell at 980°. Draper heated a strip of platinum with a battery in a dark room, waited until his eyes were sensitive to faint light, and watched for the first glow. He set "the temperature of incandescence" at 977 °F, which is 525 °C, and argued that different bodies begin to glow at the same temperature. He added a caution that applies to every hot shop: "It is to be understood, of course, that this is in a very dark room."
So 525 °C is where the glass's own light disappears even for a dark-adapted eye, and in a lit studio, by Draper's caution, it fades from view sooner. For soda-lime glass, 525 °C is also close to where the material loses the ability to correct its own mistakes.
Glass cooled unevenly locks internal stresses in, and the Lehigh lecture gives two reference points for getting them out. At the annealing point, a viscosity of 10 to the power 12.4 pascal-seconds, "the internal thermal stresses present in the glass are relieved by viscous relaxation within 15 minutes." At the strain point, 10 to the power 13.5, relieving them "may last about 15 hours," and below it, it is "practically impossible." A technical note on Corning's code 0215 soda-lime glass lists its annealing point at 545 °C and its strain point at 511 °C. On Zanotto's curve the annealing point falls at about 526 °C.
Draper's 525 °C sits inside that band. As soda-lime glass cools through the last temperatures at which it can still relax, the light it gives off goes out. Nothing in physics ties the two together. The glow, by Draper's account, depends on temperature alone, while the relaxation depends on the recipe of the glass. For this glass they happen to land in the same place, and the timing could hardly be worse: the glassblower's most direct instrument goes dark at the moment it would matter most.
So glassmaking does not ask anyone to see in the dark. Glass articles, the Lehigh lecture says, are brought "to a temperature-controlled kiln, or Lehr, for annealing," and between the annealing point and the strain point they "should be cooled down gradually, slowly and uniformly." The most delicate stage of a piece's life is decided in advance and handed to a machine running a schedule, because that stage falls exactly where the people in the room can no longer watch it.
Two habits from the hot shop carry straight into running software.
One is to watch the slope, not the level. Many of the systems engineers run have the glass's shape: they get worse faster the closer they are to a limit. In the simplest queueing model, a single server with random arrivals and random service times, the average number of jobs waiting or in service is ρ/(1−ρ), where ρ is the fraction of time the server is busy. Going from 50 to 55 percent busy moves that number from 1 to about 1.2. Going from 90 to 95 percent moves it from 9 to 19. A dashboard reading of 95 percent reports the level. The operator who has worked near the wall reads the slope, and acts while the margin can still buy time.
The other is the annealer itself. Find the stage where your instruments go dark: the point in a deploy, a migration, a failover or an incident where nobody can see what the system is doing until it is too late to act. Then decide, before you get there, what slow and automatic thing will happen in that band. It might be a timeout that rolls back on its own, a canary that drains traffic without waiting for a person, or a rate limit that engages before saturation rather than after it. Below 525 °C, glassmaking stops trying to see and puts the glass in a kiln on a schedule.
O'Connor's beginners could not see their problem without already knowing the answer. An annealer is how a craft makes sure that, at the one stage nobody can see, the answer is already in place.
Sources
How this was checked. Every quotation was checked word for word against the saved text of the source named beside it; Draper's paper was read in the scanned 1847 volume, and the exclamation marks in the "cookie" quotation are as O'Connor's field notes print them. The doubling intervals, the effect of a 20-degree drop, the temperatures of the named points on Zanotto's curve (computed from his Table I constants and the lecture's viscosity definitions) and the queueing figures were computed for this essay; none of them is printed in a source. As a check on the method, Zanotto's own equations and constants for a medieval glass, run the same way, give 773 years at 414 °C against his stated 800.
Chain of Consciousness keeps a tamper-evident, dated record in hash-linked entries, so what your systems did during a rollback, a failover or a drained canary can be checked afterward, even though nobody was watching at the time.
pip install chain-of-consciousness npm install chain-of-consciousness
Or start without installing anything: Hosted Chain of Consciousness.