It’s the late 90s and I was working in IT for a small (200 employees) company.
One of my tasks was performing daily data backups, which ran for about four hours and started after everyone else had left for the day.
We had an intercom system in the building, where if you picked up your phone and dialed a specific extension, you could page someone, or make announcements so everyone could hear.
One night I was in the process of running Ethernet cable from switches on the second floor down to individual terminals on the first floor, so there was a lot walking around the building.
I like to have music when I’m working, but I didn’t have an MP3 player at the time, so I thought… “I have music on my PC and no one is around. Why don’t I just…”, and I went around and cranked the volume knob on every ceiling intercom speaker in the whole building. Then I dialed into the intercom from the phone in my office, put the handset up as close as I could to my PC speakers, and boom, music throughout the building.
Cut to a few hours later, it’s late, I’m tired, the backup is finished, so I close down the system, hit the lights, lock the doors, and go home.
The next day, I come in, not thinking about anything from the night before, and get to work. About 11:00am, the owner’s secretary got on the intercom to page someone… and blew everyone out of their seats. There was an almost collective, “What the fuck!!!” from everyone, and an “Ohhh, shiiit” from me, because of course I immediately knew what happened.
I ran around like a madman, turning the volume down on all of the speakers that I had forgotten the night before, while saying, “No idea… Yeah it’s crazy… No no, it’s fixed now…”
No one ever found out exactly why the volume was so loud that morning.


Tooks out search on Wikipedia. I exposed a path that can perform an unconstrained Powerset Construction concurrently on 7 servers. It filled up memory and died. User retried. While cluster OOMing.
ELI-non-IT this “oh shit” moment?
Wikipedia has a search feature. Not many people use it, but Wikipedia is huge, so enough people do. Back then, a couple dozen full text searches a second. And maybe 100 quick prefix searches.
Anyway, one thing I made for it was a way to search the wiki text with regular expressions. It’s useful for maintainers.
Wikitext is the funny markup language Wikipedia uses. If you know html or md it’s like that, but it can kind of “call” other pages. Whatever. Doesn’t matter. It’s text. And people have strong opinions about it.
You can see it by clicking “edit” (sometimes looks like a pencil) and then clicking “edit without logging in” then “source editing”. Those lovely boxes at the top for a person are made with something like
{{Infobox person | name = Douglas Adams | birth_place = [[Cambridge]], England }}Imagine you have 1000 articles that use
birthplaceinstead ofbirth_place. Both work, but it’s driving you and all your editor friends crazy there are two ways to do it. This is believable. Trust me.You want to find every page that has an
{{Infobox personfollowed bybirthplace. Then fix them. The nerds who came before us invented a language to ask that question called “regular expressions”. The one for this looks like\{\{Infobox\wperson.+birthplace. Maybe. Nerds will know regular expressions are bad for this. But nerds will also know that regular expressions being bad has never stopped anyone from using them anyway.One way to run these regular expressions is to convert them into an https://en.wikipedia.org/wiki/Nondeterministic_finite_automaton . You build these in memory and they are not big. But you can’t really run them directly. They are pretty and fairly easy to read once you get used to them. But you can’t easily ask “does this match this text”. At least, not with the tools I had.
I could only run https://en.wikipedia.org/wiki/Deterministic_finite_automaton . It’s deterministic! Much nicer. And you can go from a nondeterministic one to a deterministic one. Easy. Our forenerds solved the problem. You use a https://en.wikipedia.org/wiki/Powerset_construction .
The trouble is, it can make very very very big deterministic finite automata. Like, if there are 3 states in the nfa you can get and 8 state dfa. 4 is 16. 5 is 32. In the worst case. Usually you get much better. But a fairly big but not super frightening regular expression can turn into a big nfa. And the dfa would need more states then there are grains of sand on earth. Too big for computer.
So, user asked for this regular expression. The search servers tried to convert it, and filled up their memory and died. They ran Out Of Memory. OOM. The usual thing is you copy the full memory so you can look later and start the server again.
But that can take time. And sometimes people are confused and haven’t set up the restart to be automatic. And they can’t find you. So search stays broken for a while.