4 ms·
So one flight gets canceled. There doesn't exist an automated way for the crew to log in the system that the flight is canceled. The crew of the cancelled fli
by nwallin 4y ago
So one flight gets canceled.
There doesn't exist an automated way for the crew to log in the system that the flight is canceled.
The crew of the cancelled flight calls in to a call center in a central office. There's a person there on the other end. The crew tells them that SW1234 or whatever was canceled. The person manually enters into SkySolver that the flight was canceled. SkySolver runs the algorithms or whatever and reschedules/reshuffles planes and crews to ensure that all passengers get to where they're going. SkySolver automatically sends out the corrections to all the crews and agents what the new crew/aircraft/flight assignments are.
This works great if the number of canceled flights is less than the capacity of the call center. But if the number of canceled flights exceeds the capacity of the call center by 1 aircraft, SkySolver does not know that the canceled flight did not arrive at its destination, and is unable to perform its next leg. It does not know that there's still an aircraft taking ramp space, so when/if the next flight arrives they have nowhere to park. So your 1 cancelled flight turns into 3 cancelled flights. But we're already over the capacity of the call center; so those 3 cancelled flights turns into 9 cancelled flights. 27 cancelled flights. 81 cancelled flights. 243 cancelled flights.
All cancelled flights.
- mrandish 4y agoYep, my guess is that the key point of failure tipping the system into a runaway cascade was increased latency in ingesting status reports on location/availability of crew. There probably exists an unknown "delay time threshold" in which to process and act on changing status updates. And once the average status update time exceeds that latency it just snowballs because the internal model the system is solving for is sufficiently out of sync with the real world that its "forward fixes" are causing new future failures approaching a 1:1 rate so the system can never catch up.
- nlstitch 4y agoIs there any way you think the software could have corrected this?