The Museum of Software Disasters

Rockets, fortunes, and in a few cases lives, lost to a handful of lines of code. Walk the hall and see exactly what went wrong.

By Bogdan Sandu for TMS Outsource

Enter the hall

Every exhibit here is a real failure, documented by the inquiry, court case or postmortem that followed it.

Some cost money. Some cost missions. A few cost lives.

Almost all of them came down to an assumption nobody wrote down.

Therac-25

June 1985 to January 1987. Four clinics in the United States and Canada.

What failed
Control software of a radiation therapy machine built by AECL
The bug
A race between the operator's keyboard and the routine that set up the beam
Harm
Six massive overdoses. At least three patients later died of causes attributed to them.
Outcome
Recalled in 1987 and fitted with interlocks that do not depend on the software

A machine that trusted its own software

The Therac-25 could treat a patient with a low-current electron beam or with X-rays. To make X-rays, the beam current is about 100 times higher, and a metal target and a flattening filter, mounted on a turntable, have to sit in the beam path. Eleven machines were installed, five in the US and six in Canada.

Its predecessors, the Therac-6 and Therac-20, had independent hardware interlocks that stopped the beam if anything was out of place. For the Therac-25, AECL let the software do most of that checking. Parts of the software were carried over from the older machines, where the hardware had been catching problems the software never saw.

The console

In Tyler, Texas, in March and April 1986, an experienced operator typed a prescription quickly, entered X for the mode out of habit, and corrected it to E with the cursor keys. Try the same thing.

Prescription: electron beam, 22 MeV, 180 rad


        

Bending magnets idle

Beam parameters
-
Turntable position
-

The screen is simplified. The sequence, the 8-second magnet window, Malfunction 54 and the dose monitor reading follow Leveson and Turner's investigation of the accidents.

What the software did

The mode and energy the operator typed were stored in one shared two-byte variable. One task read the high byte to load the beam parameters. Another read the low byte to move the turntable.

After data entry, a routine loaded the beam parameters and spent about 8 seconds setting the bending magnets. It looked for edits only on its first pass through a delay loop. An edit made after that changed the screen and moved the turntable, but the beam parameters were never reloaded, and nothing compared the two.

The hospital's physicist reproduced the fault only once he typed as fast as the operator. Data-entry speed was the key factor, and AECL could not reproduce it until he told them.

The last accident, in Yakima, Washington, in January 1987, came from a different bug: a one-byte counter that rolled over to zero on every 256th pass, which skipped a collimator check at the moment the operator gave the set command.

The record

  1. First known overdose, in Marietta, Georgia. Its cause was never established.
  2. First overdose in Tyler, Texas. The console shows Malfunction 54, twice.
  3. Second Tyler overdose. Within days, the physicist and the operator reproduce the fault.
  4. The US Food and Drug Administration declares the Therac-25 defective.
  5. Second overdose in Yakima, from the counter that rolls over to zero.
  6. FDA Notice of Adverse Findings. Routine use stops until August 1987.

Why nobody caught it

The error messages said things like Malfunction 54, which the manual did not explain. Treatment pauses happened often and were almost always harmless, so pressing P to proceed was routine.

One programmer wrote the software, in assembly, with little documentation and mostly system-level testing. A 1983 safety analysis left software errors out and put the chance of the computer selecting the wrong mode at 4 in a billion.

After each reported injury, AECL told the clinic that the machine could not have caused it.

When software takes over a safety check that hardware used to do, a failure in the code stops being caught by anything else. Keep the interlocks.

Sources

Nancy G. Leveson and Clark S. Turner, An Investigation of the Therac-25 Accidents. IEEE Computer, vol. 26, no. 7, July 1993, pages 18 to 41.

Deaths: Leveson and Turner describe deaths and serious injuries without giving a total. Two patients in Tyler died of the overdoses. A patient in Yakima, already terminally ill, died of complications related to the overdose.

Four years later, one switch in New York set off a failure that spread across AT&T's long-distance network.

AT&T, 15 January 1990

About 2:30 pm to 11:30 pm Eastern. The long-distance telephone network of the United States.

What failed
Recovery software in AT&T's 114 4ESS long-distance switches
The bug
Code that mishandled two messages arriving less than a hundredth of a second apart
Lost
Roughly 50 to 70 million calls failed to connect, by press estimates
Cost
More than $60 million in lost revenue is the figure usually quoted. AT&T published none.
Switches down
0 of 114
Resets
0
Signaling load
Normal

An illustrative model: 114 switches, with the links and timing simplified.

Nine hours of chain reaction

In mid-December 1989, AT&T loaded new software into all its 4ESS switches, so they could react faster when the signaling network had trouble. It had been tested in the lab.

At about 2:30 pm on 15 January, a switch in New York had a minor hardware fault. It told the switches connected to it not to send calls, spent 4 to 6 seconds recovering, then came back and started sending call attempts again.

A neighbour that received the first attempt began updating its records to mark New York as back in service. For several seconds, the processor doing that, the Direct Link Node, was vulnerable. If a second attempt arrived within a hundredth of a second, data was damaged and the processor took itself out of service. Its duplicate took over, received the next close pair and failed the same way.

Now the neighbour was cut off. It recovered, came back, sent its own call attempts, and passed the failure on. Under the afternoon's traffic, the switches kept knocking each other out.

The misplaced break

AT&T called the cause a software flaw and never published the code. The most repeated explanation, from later accounts, is a C break statement meant to leave an if that left the whole switch instead, so the code skipped ahead and overwrote data it was still using.

switch (message) {
case INCOMING_MESSAGE:
    if (sender_out_of_service) {
        if (write_buffer_empty) {
            mark_in_service(sender);
        } else {
            break;
        }
    }
    process_message();
    break;
}
do_optional_parameter_work();

An illustration of the pattern described by Bruce Sterling (1992) and Dennis Burke (1995). It is not AT&T source code.

How it stopped

Standard restoration procedures did not work. Engineers at AT&T's network operations centre in Bedminster, New Jersey, and at Bell Labs cut the message load by suspending signaling on the backup links. The last link was cleared at 11:30 pm.

The next day AT&T took the update out of the switches, went back to the previous program, reproduced the failure in the lab and corrected the code.

Recovery code runs only after something else has gone wrong, which makes it the code that gets exercised least. On 15 January it ran over and over, in switch after switch.

Sources

AT&T, Technical background on AT&T's network slowdown, January 15, 1990. Reproduced in RISKS Digest, vol. 9, issue 63, 29 January 1990.

Cause of AT&T network failure, quoting Telephony, 22 January 1990, page 11. RISKS Digest, vol. 9, issue 62, 26 January 1990.

Bruce Sterling, The Hacker Crackdown, part one, 1992.

Dennis Burke, All Circuits are Busy Now: The 1990 AT&T Long Distance Network Collapse. California Polytechnic State University, 1995.

Call counts and cost come from the press and later accounts, not from AT&T, and vary between them.

A year later, a clock that had drifted by a third of a second left a barracks in Dhahran undefended.

Patriot missile, Dhahran

25 February 1991, about 8:30 pm local time. The Gulf War.

What failed
The time calculation in the Patriot's weapons control computer
The bug
Time in tenths of a second, multiplied by a value of 1/10 chopped to 24 bits
Lost
An incoming Scud was never tracked. It hit a US Army barracks and killed 28 soldiers.
Uptime
The battery had been running for about 100 hours without a restart

A tenth of a second, almost

The Patriot's computer counted time since it was switched on in tenths of a second, as a whole number. To use that time, the software multiplied the count by 1/10. In binary, 1/10 never ends, and the register kept only 24 bits, so the stored value was a fraction too small.

Each tick lost about 0.000000095 seconds. Over an hour that is nothing. After 100 hours it is 0.34 seconds, and a Scud travels more than half a kilometre in that time.

The system used the time to predict where a detected missile would be next, and looked only inside that window, the range gate. With the clock off, the gate was in the wrong place. The Scud was not inside it, and the system decided there was no target.

The drift lab

Drag the uptime and watch the range gate slide away from the missile. The figures follow the reconstruction in the GAO's report.

Real time
0
Time the computer calculates
0
Clock error
0
Range gate shift
0

1/10 chopped to 24 bits = 0.099999904632568359375

The shift uses about 2,000 metres per second, the speed implied by the GAO's table. At the often quoted Scud speed of 1,676 m/s, 100 hours gives about 575 m. The gate size here is illustrative, set so the target leaves it after about 20 hours, as the GAO table notes.

Two weeks of warnings

  1. Israeli data reaches the Patriot Project Office: after 8 hours of continuous running, the range gate shifts by 20 percent.
  2. A software update that compensates for the error is released.
  3. Users are warned that very long run times can shift the range gate. The message does not say how long is too long.
  4. About 8:30 pm. Alpha Battery, running for about 100 hours, does not track the incoming Scud.
  5. The updated software arrives in Dhahran.

Why nobody caught it

The Patriot was designed to run for a few hours at a time in one place, against targets up to about Mach 2. In the Gulf it ran for days against Scuds flying at about Mach 5. Both made the drift matter more.

The software had already been partly corrected. A more precise time conversion had been added, but not every place that needed it called it. When one time in a calculation used the new routine and the other used the old one, the two errors stopped cancelling out.

Whether a Patriot would have destroyed this Scud with a correct clock cannot be known. The system's overall record against Scuds in the war is itself disputed.

An error that grows with time needs a stated limit. Here the limit lived only in an assumption about how long anyone would leave the system running.

Sources

US General Accounting Office, Patriot Missile Defense: Software Problem Led to System Failure at Dhahran, Saudi Arabia. GAO/IMTEC-92-26, February 1992.

Robert D. Skeel, Roundoff Error and the Patriot Missile. SIAM News, vol. 25, no. 4, July 1992.

Time of day: the GAO report gives none. About 8:30 pm is the time most sources give.

Five years later, in French Guiana, a rocket was destroyed by software it did not need.

Ariane 5, Flight 501

4 June 1996. Kourou, French Guiana. 37 seconds of flight.

What failed
Inertial reference software reused from Ariane 4
The bug
A 64-bit float converted to a 16-bit signed integer
Lost
The rocket on its maiden flight and the four Cluster science satellites it carried
Cost
About $370 million is the most cited estimate. The inquiry report gives no figure.
Try the overflow yourself

Forty seconds of a job that no longer existed

Before lift-off, the rocket's inertial reference system (SRI) runs an alignment routine that works out exactly where the platform is pointing. On Ariane 4, that routine kept running for about 40 seconds after lift-off. It let engineers restart a countdown after a late hold without waiting 45 minutes or more for a fresh alignment. That feature was used once, on Flight 33 in 1989.

Ariane 5 had no use for it. The software was kept anyway, for commonality, and it kept running into the flight.

One of its outputs, a value called BH (horizontal bias), grows with the horizontal speed of the rocket. Ariane 5 builds horizontal velocity about five times faster than Ariane 4 in the first part of its flight. BH grew past what the code had been written to expect.

The overflow lab

The flight software stored BH as a 64-bit floating point number, then converted it to a 16-bit signed integer, which can hold whole numbers from −32,768 to 32,767. Drag the flight clock and watch the two rockets.

Ariane 4 trajectory

0

Ariane 5 trajectory

0

16-bit integer: fits. Both computers healthy.

BH_Int := Integer_16 (0.0);

The numbers in this lab are illustrative: the inquiry report does not publish BH values, so the curves are scaled so that Ariane 5 crosses the limit at H0 + 36.7 seconds, as it did on the day, and Ariane 4 grows five times slower. The conversion and the 16-bit limit are real.

How the flight ended

  1. 09:33:59 local time. Main engine and both boosters ignite. Lift-off is nominal.
  2. The back-up SRI stops with an Operand Error in the BH conversion.
  3. The active SRI hits the same error in the next data cycle and shuts down too. Same hardware, same software, same input.
  4. The failed SRI sends a diagnostic bit pattern. The on-board computer reads it as flight data and swings the booster and main engine nozzles to full deflection to correct an attitude error that never happened.
  5. At an angle of attack above 20 degrees, aerodynamic loads tear the boosters off the core stage. The self-destruct system fires, about 4 km up.

Why nobody caught it

The developers had analysed every float-to-integer conversion that could overflow. They found seven risky variables, protected four, and left three unprotected, BH among them. Part of the reason was an 80% maximum workload target for the SRI computer. No justification was written in the source code.

The assumption that BH could never get that large was true for Ariane 4. It was never checked against Ariane 5, because the project partners had agreed not to include Ariane 5 trajectory data in the SRI requirements and specification.

And the redundancy did not help. Two identical computers running identical code on identical data fail identically.

Reused code carries the assumptions of the system it was written for. When the system changes, those assumptions need to be found, written down and tested again.

Sources

ARIANE 5 Flight 501 Failure, Report by the Inquiry Board, chaired by Prof. J. L. Lions. European Space Agency and CNES, Paris, 19 July 1996.

Cost: estimates for the rocket and payload vary between sources; $370 million is the figure most often quoted.

Three years later, two teams on the same mission measured the same push in different units.

Mars Climate Orbiter

23 September 1999, 09:04:52 UTC. Last signal, 49 seconds early, as the orbiter passed behind Mars.

What failed
A ground software file used to model the orbiter's small thruster firings
The bug
Data delivered in pound-force seconds, read as newton-seconds. Off by a factor of 4.45.
Lost
The orbiter, after a 9.5-month cruise. It reached Mars about 170 km lower than planned.
Cost
About $327.6 million for the program that flew it and the Mars Polar Lander, launches included

Small forces, every day

The orbiter used reaction wheels to hold its attitude. Every so often, small thruster firings unloaded the momentum they built up. Each firing nudged the trajectory a little, so the details were sent to Earth, processed by a ground program called SM_FORCES, and handed to the navigators at JPL.

The interface specification said that file had to be in newton-seconds. SM_FORCES wrote it in pound-force seconds. The navigators read it as metric, so every firing looked 4.45 times weaker than it was.

The orbiter's solar array was on one side only, so sunlight pushed it unevenly, and the firings happened 10 to 14 times more often than the navigators expected. Most of the thrust also pointed across the line of sight to Earth, where Doppler tracking could barely see it.

The units lab

One thruster firing, as it travelled from the spacecraft team to the navigators. Pick the units the file was written in.

Written in the file
1.00 lbf·s
Read by navigation
1.00 N·s
Actual impulse
4.45 N·s
Closest approach to Mars
about 57 km

One pound-force is 4.448 newtons. The altitudes are the investigation board's: 226 km planned, about 57 km reconstructed afterwards, 80 km the lowest the orbiter could survive. The board gives no simple formula linking the two, and this lab does not invent one.

Nine months of drift

  1. Launch from Cape Canaveral on a Delta II.
  2. After four months of file format errors, the thruster data becomes usable. Within a week it looks wrong. The concerns travel by e-mail.
  3. A course correction aims for a 226 km closest approach. Within a week, estimates drop to 150 to 170 km.
  4. An hour before arrival, the estimate is as low as 110 km. An extra correction burn is discussed and never made. Contact is lost at 09:04:52 UTC.
  5. Navigators find the factor of 4.45.
  6. The investigation board publishes its report.

Why nobody caught it

The navigation team had joined shortly before launch and had not been part of the design reviews or the ground software testing. Discrepancies between navigation solutions were noticed and discussed, but the team relied on e-mail instead of the formal incident and anomaly reporting process, and the problem slipped through.

No one checked that the file followed its specification. The emergency correction burn had never been planned or tested in enough detail to be used with confidence.

"The problem here was not the error," said NASA's Edward Weiler, but the failure of the checks meant to catch it.

A number without a unit is a guess. Put the units in the data, check them at every interface, and treat a worry in an e-mail as a bug report.

Sources

NASA, Mars Climate Orbiter Mishap Investigation Board Phase I Report, 10 November 1999.

NASA, Report on Project Management in NASA by the Mars Climate Orbiter Mishap Investigation Board, 13 March 2000.

NASA Release 99-113, Mars Climate Orbiter Team Finds Likely Cause of Loss, 30 September 1999.

NASA/JPL, Mars Climate Orbiter Arrival press kit, September 1999, and NSSDC Master Catalog entry 1998-073A.

Cost: NASA does not split the $327.6 million between the orbiter and the lander. Altitudes: on the day of the loss, JPL quoted 150 km planned, about 60 km actual and 85 km as the survivable minimum. The board's later figures are used here.

Three months later, the whole world waited for midnight.

Y2K

Midnight, 31 December 1999. Every computer that stored the year in two digits.

What failed
Programs that stored 1999 as 99, and so read 2000 as 1900
The bug
Two digits saved precious memory in the 1960s. The programs outlived the saving.
Cost
About $100 billion to fix in the US alone. Worldwide estimates run from $300 billion to over $800 billion.
Outcome
Mostly prevented. The failures that got through were small, and a few were not.

Two digits of savings

In the 1960s, data lived on 80-column punched cards and memory cost thousands of dollars per kilobyte. Writing 64 instead of 1964 was a sensible saving, and programmers did not expect their code to be running 30 years later. It was, and the convention spread into file formats, databases and chips.

Bob Bemer published a warning in 1971. Peter de Jager's article Doomsday 2000, in Computerworld in September 1993, is usually credited with making people listen.

The rollover lab

Four ordinary calculations, each written with two-digit years. Roll the clock past midnight, then switch to four digits.

System clock 31-12-99 23:59:55

Age of someone born in 1964
35
Card that expires in January 2000
Valid
Records sorted by year
97, 98, 99
Year printed as "19" + years since 1900
1999

Each row matches a documented failure: ages in a hospital screening system, card terminals rejecting cards that expired in 2000, records sorting in the wrong order, and websites that printed the year as 19100.

What actually broke

  1. Cards and passes that expire in 2000 start getting rejected. Visa asks banks to stop issuing cards with a 00 expiry date.
  2. Card terminals in the UK reject valid cards, because they look a few working days ahead and land in 2000.
  3. A US spy satellite ground station stops processing data for hours, and runs at reduced capacity for about three days. The cause is a faulty Y2K fix applied just before New Year.
  4. Websites, including the US Naval Observatory's, show the date as 19100. Japanese nuclear plants report minor glitches in monitoring systems, none affecting safety.
  5. A second test: 2000 is a leap year. Some systems in Japan, New Zealand and Louisiana fail to recognise the date.
  6. A hospital in Sheffield reveals that its screening software miscalculated mothers' ages from January to May 2000. More than 150 women were wrongly told their risk of a Down's syndrome pregnancy was low.

A disaster that mostly did not happen

The US Senate committee on the problem concluded that the effort was justified and the money necessary. Among US federal systems critical to their missions, the share that handled 2000 correctly rose from 21 percent in May 1997 to 99.9 percent by December 1999.

Critics point out that some countries spent far less and had no more trouble, though they were also less computerised, and that some of the money, especially at small organisations, was probably wasted. Nobody can prove what would have happened without the work. What is clear is that the large, date-dependent systems were found, fixed and tested, and the failures that got through were small.

The next date is already known. Systems that count seconds since 1970 in a signed 32-bit number run out at 03:14:07 UTC on 19 January 2038, and jump back to December 1901.

Every shortcut has an expiry date. The cheapest time to write it down is when you take the shortcut.

Sources

US Senate Special Committee on the Year 2000 Technology Problem, Y2K Aftermath: Crisis Averted. Final Committee Report, 29 February 2000.

US General Accounting Office, Year 2000 Computing Challenge: Lessons Learned Can Be Applied to Other Management Challenges. GAO/AIMD-00-290, September 2000.

US Department of Commerce, The Economics of Y2K and the Impact on the United States, November 1999.

Martyn Thomas, What Really Happened in Y2K? Gresham College lecture, 4 April 2017.

Steven Lee Myers, Y2K: In One of Few Problems, Link to Spy Satellite Fails. The New York Times, 2 January 2000.

Down's Syndrome screening failures linked to Y2K bug. The Register, 14 September 2001, based on BBC News reporting.

Spending: nobody knows the true worldwide total. The figures above are estimates made at the time.

Twelve years later, a trading firm lost more than $460 million before most people had finished their morning coffee.

Knight Capital

1 August 2012, from 9:30 am Eastern. About 45 minutes.

What failed
SMARS, the router that sent Knight's orders to the New York Stock Exchange
The bug
New code reused an old flag, and one of eight servers still ran the code that flag used to switch on
Orders
More than 4 million executions in 154 stocks, over 397 million shares
Lost
More than $460 million. Knight needed a $400 million rescue within a week.

A flag with a past

Knight was preparing for the NYSE's Retail Liquidity Program, which started on 1 August 2012. The new code went into SMARS, and it reused a flag that had once activated a feature called Power Peg.

Power Peg had not been used since 2003, but its code was still there. In 2005, the part that counted how much of an order had already been filled was moved elsewhere, and Power Peg was not tested again. If it ever ran, it would keep sending orders without noticing they had been filled.

From 27 July, a technician copied the new code to the eight SMARS servers by hand. One server was missed. No second technician checked the work, and Knight had no written procedure that required it.

The deployment

Deploy the code, open the market, and then do what Knight's team did when they tried to fix it.

Clock
9:30 am
Executions
0
Shares traded
0
Unwanted positions
$0

The counters spread the totals in the SEC's order evenly over the 45 minutes. The real pace is not published.

The morning

  1. The new code starts going onto the SMARS servers, one by one. The eighth is never updated.
  2. SMARS starts sending automated emails that mention Power Peg. There are 97 of them before the open. Nobody acts on them.
  3. The market opens. 212 incoming orders reach the eighth server, and Power Peg starts firing child orders.
  4. Trying to stop it, the team removes the new code from the seven good servers. All eight now run Power Peg.
  5. After about 45 minutes, the orders stop.
  6. A group of investors puts in $400 million. In December, Knight agrees to merge with GETCO.

Why nobody caught it

The deployment was manual, unchecked and undocumented. Reusing a flag looked harmless because the code it used to switch on was considered dead. Nobody had removed it.

The warnings existed. The 97 emails went to a group of staff, but they were not treated as alerts, and nothing escalated them. Knight had no automated way to stop SMARS from sending orders that far exceeded what its customers had asked for.

In 2013 the SEC fined Knight $12 million for breaking the Market Access Rule, which requires brokers to have controls that prevent exactly this.

Dead code is not dead while something can still call it. Delete it, and never give an old switch a new meaning.

Sources

US Securities and Exchange Commission, In the Matter of Knight Capital Americas LLC. Administrative Proceeding File No. 3-15570, Release No. 70694, 16 October 2013.

SEC press release 2013-222, SEC Charges Knight Capital With Violations of Market Access Rule, 16 October 2013.

Knight Capital Group, Form 8-K filings and press releases, August 2012.

Loss: Knight's filings give figures between $440 million and $468 million depending on what is included. The SEC says more than $460 million.

A year later, a government website opened to 250,000 people at once, through a door built for far fewer.

HealthCare.gov

Opened 1 October 2013. Working for most people by 1 December.

What failed
The federal website where people in 36 US states were meant to compare and buy health insurance
The bottleneck
Everyone had to create an account first, through one identity system that could not keep up
Launch day
250,000 people at once, five times what was expected. Outages within two hours.
Cost
The main website contract grew from an estimated $58 million to $207 million

A one-lane onramp

HealthCare.gov was six large systems built by dozens of contractors, joined to five federal agencies, 36 states and more than 300 insurers. On launch day every visitor entered through the same door: an account and identity system that created a login and checked who you were before you could see a single plan. One official later called it a small, one-lane onramp to a major highway.

The warnings were in the test results. On 25 September, a test showed the site could handle about 10,000 people at once. On 27 September, a CMS official wrote that it had never handled more than 500 people filling out applications in a setup the size of launch day. In between, the hosting contractor was asked to double everything before launch.

On 1 October, 250,000 people tried at the same time. Most never got past the account step, and the account step hid deeper faults in the marketplace software behind it. The next morning, internal notes recorded six enrollments sent to insurers.

The funnel lab

Open the site on launch day and watch where people pile up. Then run the repair.

Trying to get in
0 Past the account step
0 Filling in applications
0
Clock
1 Oct, opening
Site uptime
Not measured yet

The 250,000 peak, the 10,000 and 500 test results, the uptime figures and the 50,000 target are from HHS and CMS records. The two tests measured different things on different days. The shape of the curves and the capacities between them are illustrative.

Two months to work

  1. An outside review warns of too little end-to-end testing and a launch at full volume.
  2. A test shows the site can handle about 10,000 people at once.
  3. Launch, on the same morning the federal government shuts down for 16 days. Outages begin within two hours.
  4. New leadership and a team of engineers from technology companies are brought in.
  5. Uptime rises from 42.9 percent in the first week to 95.1 percent in the last, after more than 400 fixes.
  6. The first enrollment period closes with about 8 million plan selections, 5.4 million through HealthCare.gov.

Why nobody caught it

Requirements kept changing well into 2013, until there was no time left to finish building the site and test it. The team chose to launch on the date rather than finish testing. No full end-to-end test was run; one official described the launch itself as the test.

Nobody was clearly in charge of the whole. The agency acted as its own systems integrator without the experience to do it, while each contractor tested its own part. The government's inspector general counted 18 documented warnings before launch, and the most common advice was to appoint a single lead and stop changing the requirements.

The repair worked by doing exactly that: one integrator, one command centre and daily measurement. In August 2014 the White House created the US Digital Service, building on the team that had fixed the site.

Test the whole path a user takes, at the load you expect and more. And put one person in charge of the whole.

Sources

HHS Office of Inspector General, HealthCare.gov: Case Study of CMS Management of the Federal Marketplace (OEI-06-14-00350), February 2016.

US Government Accountability Office, Healthcare.gov: Ineffective Planning and Oversight Practices Underscore the Need for Improved Contract Management (GAO-14-694), 30 July 2014.

US Government Accountability Office, Healthcare.gov: CMS Has Taken Steps to Address Problems (GAO-15-238), 4 March 2015.

House Committee on Energy and Commerce, PPACA Implementation Failures: Didn't Know or Didn't Disclose? Hearing with contractor testimony, 24 October 2013.

CCIIO war room notes, October 2013, released by the House Committee on Oversight and Government Reform.

Centers for Medicare and Medicaid Services, HealthCare.gov Progress and Performance Report, 1 December 2013.

HHS ASPE, Health Insurance Marketplace: Summary Enrollment Report for the Initial Annual Open Enrollment Period, May 2014.

The White House, Delivering a Customer-Focused Government Through Smarter IT, 11 August 2014.

Contract figures measure different things and are not the total cost of the website.

Six months later, a missing length check let anyone read the memory of a large part of the secure web.

Heartbleed

Disclosed 7 April 2014. In OpenSSL since March 2012. CVE-2014-0160.

What failed
The heartbeat feature of OpenSSL, the library that encrypted much of the web's traffic
The bug
The server trusted the length a client claimed for its message, and copied that much of its own memory back
Exposed
About 17 percent of secure web servers with trusted certificates, around half a million sites
Leaked
Up to 64 KB per request, repeatable and unlogged: passwords, cookies, private keys

Are you still there?

A heartbeat is a small message that keeps an encrypted connection alive. The client sends some bytes and says how many it sent. The server sends the same bytes back. The specification says that if the claimed length is bigger than the message, the request must be discarded.

OpenSSL's code read the claimed length, a 16-bit number of up to 65,535, and copied that many bytes starting at the message, without comparing it to what had actually arrived. Send one byte, claim 65,535, and the reply contains almost 64 KB of whatever sat next to your message in the server's memory.

The code was committed on New Year's Eve 2011 and shipped in OpenSSL 1.0.1 in March 2012. Its author later said he had simply missed validating a variable containing a length.

The heartbeat lab

Send a word to the server and say how long it is. Then lie.


        
        

The server memory here is invented and much shorter than the real 64 KB. The mechanism, the 16-bit length and the fix follow the OpenSSL source and RFC 6520.

Two years in the open

  1. The heartbeat code is committed to OpenSSL, after review.
  2. OpenSSL 1.0.1 ships with it.
  3. Neel Mehta of Google and engineers at Codenomicon find the bug independently.
  4. Public disclosure and OpenSSL 1.0.1g, which adds the missing check.
  5. CloudFlare challenges the public to steal a private key from a vulnerable test server. The first person does it the same day.
  6. The Canada Revenue Agency says Social Insurance Numbers of about 900 taxpayers were taken through the bug.

Why nobody caught it

OpenSSL secured a huge share of the internet and was maintained by a handful of people, only one of them full-time. Its foundation typically received about $2,000 a year in donations. No money was set aside for code review or audits.

The bug was a single unchecked number in C, a language that will copy past the end of a buffer if told to. It broke a rule written in the specification, and the code passed review anyway.

Within weeks, the Linux Foundation launched the Core Infrastructure Initiative, with a dozen large technology companies paying to support OpenSSL and other critical open-source projects. The OpenBSD team forked OpenSSL into LibreSSL.

Never trust a length you did not measure. And if everyone depends on a project, someone has to pay for it.

Sources

OpenSSL Security Advisory, TLS heartbeat read overrun (CVE-2014-0160), 7 April 2014.

Codenomicon, The Heartbleed Bug, heartbleed.com, April 2014.

IETF RFC 6520, Transport Layer Security (TLS) and Datagram Transport Layer Security (DTLS) Heartbeat Extension, February 2012.

OpenSSL source, commits 4817504d (31 December 2011) and 96db9023 (7 April 2014).

Paul Mutton, Half a million widely trusted websites vulnerable to Heartbleed bug. Netcraft, April 2014.

CloudFlare, The Results of the CloudFlare Challenge, 11 April 2014.

Canada Revenue Agency, statement by the Commissioner, 14 April 2014, as reported by CBC News and others.

Steve Marquess, Of Money, Responsibility, and Pride, 12 April 2014.

The same year, a pop video climbed towards a number YouTube's counter was never meant to reach.

Gangnam Style

December 2014. The near miss in this museum.

What was at risk
YouTube's view counter, which had stored counts as signed 32-bit integers
The limit
2,147,483,647, the largest number a signed 32-bit integer can hold
The video
PSY's Gangnam Style passed it in late November 2014, about 2.15 billion views by early December
Outcome
Nothing broke. YouTube had moved the counter to 64 bits months earlier.

A number with a ceiling

A signed 32-bit integer has 32 binary digits. The first one marks the sign, which leaves 31 for the value, so the biggest number it can hold is 2 to the power of 31, minus one: 2,147,483,647. Add one more and, in the way almost all computers store integers, the sign bit flips and the number becomes minus 2,147,483,648.

When YouTube's counters were designed, that ceiling looked unreachable. Then a Korean pop video uploaded on 15 July 2012 became the first to reach a billion views that December, and two billion on 31 May 2014.

On 1 December 2014, YouTube posted that it had never thought a video would be watched more than 2,147,483,647 times, "but that was before we met PSY", and that it had moved to a 64-bit integer. Headlines said the video had broken YouTube. Google said nothing had actually broken. Hovering over the counter made its digits spin, a joke YouTube later called an easter egg.

The counter lab

Each square is one bit of the stored count. The red one is the sign. Take the count to the top of the 32-bit range and one view past it.

0

The wrap shown here is what a signed 32-bit integer does. YouTube's own counter never wrapped. YouTube's post gave the 64-bit limit as 9,223,372,036,854,775,808, which is one too many: the largest signed 64-bit integer ends in 807.

From a billion to the limit

  1. Gangnam Style is uploaded to YouTube.
  2. It becomes the first YouTube video to reach a billion views.
  3. The first to reach two billion.
  4. Months ahead of the limit, YouTube moves its view counter to 64 bits.
  5. The count passes 2,147,483,647. Nothing visible happens. The exact day is an estimate.
  6. YouTube announces the upgrade. Coverage says the video broke YouTube.
  7. See You Again overtakes it as the most viewed video on YouTube.

Why it did not break

Someone at YouTube watched the number climb and changed the type before it mattered. That is the whole story, and it is why this exhibit is here: most of the disasters in this hall had a moment like that which nobody acted on.

Counters that outgrow their integers are common. In 2015 the US aviation regulator ordered Boeing 787 operators to power the aircraft down periodically: a software counter inside its generator control units would overflow after 248 days of continuous power, and the units could all drop into failsafe mode at once, cutting AC electrical power. A 32-bit limit also sits under the clock at the end of this hall.

Every counter has a ceiling. Find out where yours is before your users do.

Sources

BBC News, Gangnam Style music video 'broke' YouTube view limit, 4 December 2014.

Brandon Griggs, 'Gangnam Style' breaks YouTube. CNN, 3 December 2014.

Rick Regan, Gangnam Style Video Overflows YouTube Counter. Exploring Binary, 3 December 2014.

Katie Rife, YouTube was just kidding about "Gangnam Style" breaking its view counter. The A.V. Club, 5 December 2014, citing Variety.

BBC Newsround, Gangnam Style first YouTube video to reach 2 billion views, 31 May 2014.

Andrea Dresdale, 'See You Again' breaks record for most-viewed YouTube video of all time. ABC News, 11 July 2017.

Federal Aviation Administration, Airworthiness Directive 2015-09-07, The Boeing Company Airplanes. Federal Register, 1 May 2015.

YouTube's Google+ post of 1 December 2014 is no longer online and is quoted here as reported by the BBC.

Fifteen months later, eleven lines of JavaScript disappeared, and thousands of projects stopped building.

left-pad

22 March 2016, from about 2:30 pm Pacific. About two and a half hours.

What failed
Installs of thousands of JavaScript projects from the npm registry, Babel among them
The cause
A developer unpublished 273 of his packages. One was left-pad, 11 lines long.
Reach
left-pad had been downloaded 2,486,696 times in the previous month
Outcome
npm restored the package itself, then changed its rules on unpublishing

A naming dispute

The messaging company Kik wanted the npm package name kik for its own project. The developer who already owned it, Azer Koçulu, declined to give it up. npm decided the name should go to Kik.

In response, he unpublished every package he had on npm, which the rules of the time allowed. Among them was left-pad, a function that adds characters to the start of a string until it reaches a given length.

Babel and Atom did not use left-pad directly. They used a package called line-numbers, which required exactly version 0.0.3 of left-pad. Within ten minutes someone published a new left-pad, but it had a different version number, so the builds kept failing.

The dependency lab

This is the whole package. Pull it from the registry and watch what happens to everything built on top of it.

module.exports = leftpad;
function leftpad (str, len, ch) {
  str = String(str);
  var i = -1;
  ch || (ch = ' ');
  len = len - str.length;
  while (++i < len) { str = ch + str; }
  return str;
}

your appbabel-coreline-numbersleft-pad 0.0.3


        

The code is left-pad as it was in 2016, with blank lines removed. The chain is simplified: Babel reached left-pad through line-numbers.

Two and a half hours

  1. Pacific time. Shortly after, installs start failing with a 404 for left-pad.
  2. An equivalent left-pad 1.0.0 is published. Builds still fail, because line-numbers asks for 0.0.3.
  3. npm announces it will republish the original 0.0.3 from a backup.
  4. left-pad 0.0.3 is back in the registry.
  5. New npm policy: a version can be unpublished only within 24 hours. After that, npm refuses if other packages would break.

Why it could happen

The registry let anyone remove a package that thousands of others depended on, at any time, with no warning to the people downstream. And a large share of modern JavaScript sat on chains of small packages that nobody had looked at in years.

The incident started a debate about trivial dependencies. The developer David Haney argued that functions this small should simply be written in house. Researchers later found that about 16 percent of the npm packages published before the incident were trivial in the same way. JavaScript itself gained a built-in padStart in 2017.

Every dependency is code you run but did not write, maintained by someone you have never met. Know what is at the bottom of your stack.

Sources

npm, Inc., kik, left-pad, and npm. npm blog, 23 March 2016.

npm, Inc., changes to npm's unpublish policy. npm blog, 29 March 2016.

Azer Koçulu, I've Just Liberated My Modules, 23 March 2016.

Chris Williams, NPM left-pad chaos. The Register, 23 March 2016.

David Haney, NPM & left-pad: Have We Forgotten How To Program? 23 March 2016.

Rabe Abdalkareem and others, Why do developers use trivial packages? An empirical case study on npm. ESEC/FSE 2017, extended in Empirical Software Engineering, 2020.

No source counts the builds that failed. npm described the damage as many thousands of projects.

Five years later, one command erased Facebook, Instagram and WhatsApp from the internet.

Facebook goes dark

4 October 2021, from about 15:39 UTC. Nearly six hours.

What failed
Facebook, Instagram, WhatsApp, Messenger and Oculus, worldwide, and most of the company's own internal tools
The cause
A maintenance command took down Facebook's entire backbone. The tool meant to block it had a bug.
Reach
Apps with about 3.5 billion monthly users. DNS resolvers worldwide saw 30 times their usual queries.
Not an attack
Meta found no malicious activity and no evidence that user data was compromised

Erased from the map

Two systems hold the internet together. DNS is its address book: it turns facebook.com into a numeric address. BGP is how the internet's networks tell each other which addresses they can reach. Facebook's network, AS32934, announced routes to its own DNS servers, so the rest of the internet knew where to send questions about facebook.com.

On 4 October, during routine maintenance, a command was issued to assess how much capacity the global backbone had to spare. It disconnected every link in that backbone instead, cutting Facebook's data centres off from each other and from the world. A tool that audits such commands should have stopped it. A bug in the tool let it through.

Then a safety rule made it worse. Facebook's DNS servers stop advertising their routes if they cannot reach the data centres, because that usually means the site is sick. Every site lost the backbone at the same moment, so every site withdrew. The servers kept running, but no route on the internet led to them any more.

The routing lab

Sixteen networks hold a route to Facebook's DNS servers. Run the maintenance command, look the site up, then try to get it back.


        
Clock
15:38 UTC
DNS query load
1x normal
Recovery
Idle

The times and the 30x query load come from Cloudflare's measurements. The networks, their numbers, the speed of the withdrawal and the recovery steps are illustrative: real withdrawals spread in seconds, and Meta did not publish the command or the bug.

Locked out

  1. A spike of routing updates from Facebook's network. Traffic to its apps falls away.
  2. facebook.com stops resolving on Cloudflare's public resolver. Lookups return SERVFAIL.
  3. Facebook is no longer announcing any route to its DNS servers.
  4. Remote access is down with the network. Engineers go to the data centres, where strict physical security takes extra time.
  5. Facebook's routes start coming back.
  6. Cloudflare sees Facebook reconnected and DNS working. The apps take longer to recover fully.

Why nobody caught it

The last line of defence was a piece of software, and it had a bug. Behind it, a rule that protects one failing site turned a backbone failure into a total disappearance, because every site obeyed it at once.

Recovery depended on the systems that were down. The tools engineers use to investigate outages ran on Facebook's own network and DNS, and so did the remote access to the data centres. Reports at the time said some staff badges stopped working; Meta has said only that physical security slowed engineers getting in. Once the backbone was back, power use in individual data centres had dropped by tens of megawatts, and switching everything on at once risked a new round of failures. Meta had rehearsed this kind of restart in drills it calls storms, but never one that took out the whole backbone.

Meta said it would simulate a backbone loss in future drills and strengthen its testing. It did not publish the command, the bug or a fix. Estimates of the revenue lost range from about $3.3 million in US advertising to about $100 million, depending on what is counted.

Never let the tools you need to fix an outage depend on the system that is out. And rehearse losing everything, not just a part.

Sources

Santosh Janardhan, Update about the October 4th outage. Engineering at Meta, 4 October 2021.

Santosh Janardhan, More details about the October 4 outage. Engineering at Meta, 5 October 2021.

Celso Martinho and Tom Strickx, Understanding How Facebook Disappeared from the Internet. Cloudflare, 4 October 2021.

Doug Madory, Facebook's historic outage, explained. Kentik, 5 October 2021.

Salvador Rodriguez, Facebook shares drop nearly 5% after major site outage and whistleblower interview. CNBC, 4 October 2021.

Tom Huddleston Jr., Facebook's outage cost it an estimated $100 million in lost revenue. CNBC, 6 October 2021.

Brian Krebs, What Happened to Facebook, Instagram, & WhatsApp? KrebsOnSecurity, 4 October 2021.

Times are as observed by Cloudflare and Kentik. Meta published no clock times. Dollar figures are outside estimates.

Three years later, one configuration file stopped millions of computers from starting at all.

CrowdStrike

19 July 2024, 04:09 to 05:27 UTC. 78 minutes of a bad file, days of recovery.

What failed
Windows computers running CrowdStrike's Falcon security sensor
The bug
A content file used a 21st input field. The sensor code only ever supplied 20.
Reach
About 8.5 million Windows devices, by Microsoft's estimate: airlines, hospitals, banks, broadcasters, emergency lines
Cost
Estimated at $5.4 billion for the US Fortune 500 alone, Microsoft excluded

Configuration, not code

Falcon runs a driver deep inside Windows, loaded early in boot, so it can see attacks as they happen. CrowdStrike updates its detection rules through small content files, called channel files, which it can push to every customer within minutes. They are configuration data, not code, and they skipped the slower release process that code went through.

In February 2024, a sensor update added a new template for watching how Windows programs talk to each other. The template defined 21 input fields. The code that fed it supplied 20. For five months nobody noticed, because every rule shipped so far used a wildcard for the 21st field.

On 19 July, Channel File 291 added a rule that compared the 21st field against a real value. The sensor reached for an input that did not exist, read past the end of its array, and crashed Windows. Because the driver loads at boot, the machine crashed again every time it started.

The recovery lab

Push the file, withdraw it, then try to get 24 machines working again.

Crashed
0 of 24
Fixed
0
Technician
Idle

The fix steps follow CrowdStrike's and Microsoft's guidance: boot into Safe Mode or the recovery environment, enter the BitLocker key if the disk is encrypted, delete the files matching C-00000291*.sys in the CrowdStrike drivers folder, and reboot. The odds of the reboot button are illustrative.

The longest weekend

  1. Sensor 7.11 ships with the new template: 21 fields defined, 20 supplied.
  2. UTC. Channel File 291 goes out with two new rules. One uses the 21st field.
  3. The file is withdrawn. Machines that already crashed stay down.
  4. Microsoft estimates 8.5 million Windows devices were affected, under 1 percent of all Windows machines, and soon releases a recovery tool.
  5. CrowdStrike reports about 99 percent of Windows sensors back online.
  6. CrowdStrike publishes its root cause analysis.

Why nobody caught it

The tests used wildcards in the 21st field, so the missing input never mattered. The content validator checked the new rules against a template it believed had 21 inputs, which CrowdStrike's analysis calls a logic error. The interpreter that read the file had no bounds check.

And the file went to every customer at once. There was no small first group of machines to fail before everyone else did.

CrowdStrike has since added the bounds check, a check on input counts when content is compiled, tests for every field, staged rollouts that start with canary machines, and an option for customers to choose when content updates reach them. Delta Air Lines puts its own losses at more than $500 million, a figure CrowdStrike disputes.

Configuration that changes behaviour is code. Test it like code, and release it in stages, however urgent it feels.

Sources

CrowdStrike, External Technical Root Cause Analysis: Channel File 291, 6 August 2024.

CrowdStrike, Preliminary Post Incident Review and Remediation and Guidance Hub, July 2024.

David Weston, Helping our customers through the CrowdStrike outage. Microsoft, 20 July 2024.

Microsoft, New Recovery Tool to help with CrowdStrike issue impacting Windows endpoints, July 2024.

Testimony of Adam Meyers, CrowdStrike, to the US House Committee on Homeland Security, 24 September 2024.

Parametrix, CrowdStrike's Impact on the Fortune 500, 24 July 2024.

Dollar figures are estimates or claims, not audited losses.

The last exhibit began before most of the others in this hall, and it is not over yet.

Post Office Horizon

Rolled out from 1999. Convictions quashed by Act of Parliament in 2024.

What failed
Horizon, the counter and accounting system in the UK's Post Office branches, built by ICL Pathway and later Fujitsu
The bugs
Defects that made branch accounts show shortfalls that did not exist
The harm
Sub-postmasters were made to repay the shortfalls, and about 700 were convicted in Post Office prosecutions between 1999 and 2015
Redress
About £1,756 million paid to more than 14,100 claimants, as of 30 October 2026

The computer said so

Sub-postmasters run Post Office branches as independent businesses. Their contract said that deficiencies in the branch accounts must be made good without delay. When Horizon arrived, the accounts were whatever Horizon said they were. If the balance showed a shortfall, the branch had to pay it or accept it, because the system had no way to dispute a figure.

Horizon had bugs that could create shortfalls, and it did not tell users about them. Branch staff could see only a few weeks of transaction logs on till rolls, never the back-end data where a fault would show. Fujitsu staff could also change branch data remotely, in ways that looked as though the sub-postmaster had done it. The Post Office said publicly in 2015 that this was impossible. The High Court later found that statement was not true.

The Post Office was also a private prosecutor. Between 1999 and 2015 it secured about 700 convictions, for theft, fraud and false accounting, in cases where Horizon evidence may have featured. Other prosecutors brought about 280 more. Since 2000, courts in England and Wales have presumed that a computer was working correctly unless someone shows otherwise.

The counter lab

This is the Dalmellington bug, as the High Court described it in 2019. Send one pouch of cash from a core branch to its outreach branch.

What happened at the counter

    What Horizon recorded

      The mechanism, the £8,000 that became £32,000 and the 112 occurrences come from the 2019 judgment and its technical appendix. In the real case the £24,000 was corrected a few weeks later. The screens, pouch number and timings are illustrative.

      Twenty-five years

      1. Horizon begins rolling out to Post Office branches.
      2. Computer Weekly publishes the first press investigation. Sub-postmasters form a campaign group.
      3. An independent review finds two bugs that affected 76 branches.
      4. The High Court finds that Horizon's bugs could and did cause shortfalls. 555 claimants settle for £57.75 million, most of it legal costs.
      5. The Court of Appeal quashes 39 convictions, calling the prosecutions an affront to the conscience of the court.
      6. The Post Office (Horizon System) Offences Act quashes qualifying convictions in England, Wales and Northern Ireland. Scotland passes its own Act weeks later.
      7. The public inquiry publishes the first volume of its final report. Further volumes are still to come.

      Why it lasted so long

      The software was not the whole failure. In 2019 the High Court examined 29 entries in an expert's table of known problems and found that most were bugs able to affect branch accounts. The judge called the original system "not remotely robust". Some lay undiscovered for years. Others were known: a 2010 internal note on one of them listed the risk to ongoing legal cases.

      What turned defects into convictions was the way the evidence was treated. The Court of Appeal found that the Post Office presented a shortfall from an unreliable system as a proven loss and left the accused to prove otherwise, and that its failures of investigation and disclosure were so serious that the prosecutions should never have been brought.

      By June 2026 the 2024 Act had quashed 2,228 convictions belonging to 610 people in England and Wales. The public inquiry, chaired by Sir Wyn Williams, found that many hundreds of people were wrongly convicted and many thousands held responsible for losses that were illusory. In 2025 the government asked for evidence on whether the presumption that computers are reliable should change.

      A computer's output is a claim, not a fact. When people's liberty depends on it, the system has to be shown to be right.

      Sources

      High Court, Judgment (No. 6) "Horizon Issues", [2019] EWHC 3408 (QB), with Technical Appendix, 16 December 2019.

      High Court, Judgment (No. 3) "Common Issues", [2019] EWHC 606 (QB), 15 March 2019.

      Court of Appeal, Criminal Division, [2021] EWCA Crim 577, 23 April 2021.

      Post Office Horizon IT Inquiry, Final Report Volume 1, 8 July 2025, and progress update, 8 July 2026.

      Post Office (Horizon System) Offences Act 2024 and its Explanatory Notes.

      Ministry of Justice, quashed convictions management information, July 2026, and call for evidence on the use of evidence generated by software in criminal proceedings, January 2025.

      Department for Business and Trade, Post Office Horizon financial redress data as of 30 October 2026.

      Post Office Ltd, Horizon scandal: frequently asked questions, figures on prosecutions 1999 to 2015.

      Second Sight Support Services, Interim Report, 8 July 2013.

      This exhibit names no sub-postmasters. Redress totals change monthly.

      At the end of the hall there is one more clock. It is still running.

      The next one: 2038

      19 January 2038, 03:14:07 UTC.

      Many computers count time as the number of seconds since 1 January 1970. Where that count is kept in a signed 32-bit integer, it runs out at 2,147,483,647, the same ceiling the Gangnam Style counter was moved away from. One second later, the count wraps to its most negative value, and the date reads 13 December 1901.

      The clock lab

      This is the real time, stored the old way. Jump to ten seconds before the limit and watch the last second.

      0

      Until the limit
      0

      Most current operating systems and databases have already moved to 64-bit time. The risk is in what has not: embedded devices, old file formats and database columns, and code that stores time in a 32-bit field.

      The ledger

      The fourteen exhibits did not fail in the same currency. Some cost money, some cost time, some cost lives and liberty. Here is each one in its own terms, as the sources give it.

      ExhibitWhat it costThe root of it
      Therac-25, 1985Six massive overdoses; at least three deaths attributed to themA race between keyboard and beam setup, with no independent hardware interlock
      AT&T, 1990About nine hours of failed long-distance calls; more than $60 million is the usual estimateRecovery code that mishandled two messages arriving too close together
      Patriot, 199128 soldiers killedA clock that drifted the longer it ran
      Ariane 5, 1996About $370 million, the most cited estimateReused code that overflowed on new data
      Mars Climate Orbiter, 1999About $327.6 million for the programImperial units where metric was expected
      Y2K, 2000About $100 billion to fix in the US; most failures preventedYears stored in two digits
      Knight Capital, 2012More than $460 million in 45 minutesA reused flag and one server missed in a deployment
      HealthCare.gov, 2013Two months of a site most people could not useNo end-to-end test, no one in charge of the whole
      Heartbleed, 2014Private memory exposed on about 17 percent of trusted secure serversA length the server never checked
      Gangnam Style, 2014NothingA 32-bit counter, changed in time
      left-pad, 2016About two and a half hours of failed buildsEleven lines anyone could remove
      Facebook, 2021Nearly six hours offline for apps with billions of usersA bug in the tool meant to block a bad command
      CrowdStrike, 2024About 8.5 million devices down; $5.4 billion estimated for the Fortune 500An untested field in a configuration file
      Post Office Horizon, 1999 onwardAbout 700 Post Office convictions; £1,756 million in redress so farBugs, and the belief that the computer must be right

      What the hall has in common

      Almost none of these failures came from code nobody had looked at. They came from code that worked somewhere else, under other conditions: a rocket that flew slower, a clock that stayed accurate only if someone restarted it, a field that always held a wildcard, a counter that never got that high. The software did what it was written to do. The world it was written for had changed.

      And almost every one had a warning, a review or a test that pointed at the problem before it happened. The exception that proves it is the near miss in the middle of the hall, where someone watched a number climb and acted in time.

      Write down what your code assumes about the world. Then check, before every launch, whether the world still agrees.

      The museum is curated by Bogdan Sandu and published by TMS Outsource, a software development company with engineering teams in Belgrade and Bucharest.