NATS publishes preliminary report on technical incident of 8 September

aasplake about 4 hours ago 17 commentsRead Article on nats.aero

DE version is available. Content is displayed in original English for accuracy.

⚡ Community Insights

Discussion Sentiment

0% Positive

Analyzed from 669 words in the discussion.

Discussion (17 Comments)Read Original on HackerNews

zrail•about 2 hours ago

Note: NATS the UK airspace manager, not NATS the pubsub system.

verst•about 1 hour ago

And here I was surprised they would use NATS in their implementation to assign squawk codes.

macguillicuddy•about 2 hours ago

Interesting! Looks like a race condition (with a 1ms window) during squawk allocation (the process that gives each aircraft a 4 digit identifier for ATC purposes) caused data corruption. They present that window as being small but 1ms is a decent chunk of time in our work lol

crote•about 1 hour ago

More specifically: it sounds like something which was supposed to be an atomic operation was split into two by the preempting task, with corruption happening because that task also interacted with the same piece of (now-temporarily-invalid) data.

lukaslueg•about 1 hour ago

Would be interesting if the report actually got into the details. The curiously specific "approx. 1ms window" sounds like an off-the-shelf memory race condition in combination with a preempting thread, for which 1ms is an absolute eternity and Guaranteed To Happen By Tuesday This Week™. But the "database corruption" sounds like database IDs being generated from timestamps with 1ms granularity.

LtdJorge•about 1 hour ago

They might have to switch to ULID or similar.

fotta•about 2 hours ago

link to the report (pdf): https://www.nats.aero/wp-content/uploads/2026/09/NATS-Prelim...

fotta•about 2 hours ago

> The incident was triggered by a valid manual request for a squawk code. This manual request was made correctly and there was nothing abnormal or invalid about the associated flight plan.

> While this request was being processed, the NAS received a message for a higher priority activity to be undertaken which resulted in the squawk code allocation being paused while the system processed the higher priority message. Switching between different activities in response to prioritised requests is a normal function of the system; however, when the processing of the squawk allocation request resumed, the software defect meant it did not resume correctly and the resulting output was corrupted.

> The reason this scenario has not occurred before is because:

> 1. The defect existed in a specific subsection of code within a software module, with an exposure window estimated as approximately one millisecond.

> 2. For the fault to occur, a higher-priority request had to arrive during that exact millisecond while the original request was part-way through updating a value.

> 3. Had the higher-priority request arrived even one millisecond earlier or later, the update would have completed normally.

> Post-incident investigation has identified that when processing of the squawk allocation request resumed, the data associated with it had been corrupted and affected some subsequent flight data updates.

dtf•about 1 hour ago

> 3. Had the higher-priority request arrived even one millisecond earlier or later, the update would have completed normally.

Well, that's comforting to know.

BBC: "Flight chaos caused by software defect in space of a millisecond, report says"

Sky: "'Millisecond' software error caused air traffic outage that grounded thousands of flights"

The Guardian: "Flight chaos for hundreds of thousands was caused in ‘millisecond’ by software error"

Sounds like pure bad luck.

cmpxchg8b•about 1 hour ago

Safety critical systems demand formal verification. This wasn't bad luck, this was poor craft.

philipwhiuk•about 1 hour ago

Or insufficient testing?

iso1631•14 minutes ago

> 3. Had the higher-priority request arrived even one millisecond earlier or later, the update would have completed normally.

How often does that original request happen per day? How often does the higher priority activiry take?

If the original request happens 864 times a day and the high priority request ten times, there's a 1 in 25 chance it will happen in a given year.

stackghost•28 minutes ago

Very interesting. I spent a large part of my career in aerospace and never considered this failure mode before. It makes me wonder: how long is the pathological code allocation time? A few seconds, at most? We're talking about flight identification codes that are normally assigned upon takeoff and change at most a handful of times during a flight.

I assume the "manual request" is an aircraft squawking 7700 or similar, but why does the system need to interrupt an in-flight allocation in the first place? Any controllers here have insight?

One would think it would be sufficient to do something single threaded like

    if(!highPriorityQueue.empty() {
        highPriortyQueue.processOne();
    } else if(!lowPriorityQueue.empty()) {
        lowPriorityQueue.processOne();
    }

or whatever, but they're not and I'm curious why.

benmmurphy•5 minutes ago

there will be some other tasks in the system that need time guarantees around when they get scheduled for very important (safety?) reasons. so when these higher priority tasks show up the lower priority tasks get suspended.

if they could guarantee hard bounds on how long low priority tasks take to complete they could implement your scheduling algorithm. but i think in reality the low priority tasks are either not boundable or they if they do have a provable bound the bound is too high.

chrisjj•about 1 hour ago

Therac-25 called and wants its bug back.

Seriously, for 2026 this is pure amateur hour with no excuse.

imoverclocked•8 minutes ago

2026 … but the code is running from what decade and has to interact with systems from what other decades?