Hacker News Reader

Comment by speedgoose | Hacker News Reader

The scale of Uber is a bit incredible. Savings 70 000 cpu cores is a lot for a taxi app.

Calling Uber just "a taxi app" is IMHO like calling Facebook just a website. Sure, that's what the end user gets, but there is plenty of engineering involved such as maps service (I believe they have their own, but I might be wrong), geo searching for locations, batch processing for data analytics, ML experiments etc.

I believe Uber is also well known for building everything in house as opposed to using common cloud services. So they need to run their stuff.

Makes me wonder what on earth all those cpus are doing. Many people take trips by taxi, but not THAT many.

Right? Like our product isn’t Uber scale, but our Go microservices each get about 10,000 active users per core before the autoscaler adds a new instance. I’m curious their numbers.

"Uber drivers completed 4.98 billion trips in 2020, a 27 percent decrease from the 6.9 billion trips in 2019"

Now, granted, 2021 is probably down on that further but it's still a lot.

7 billion per year is 222 trips per second on average.

With each trip taking an average of 15 minutes, that’s about 200k concurrent, live sessions. Each session has at least two participants, so 400k live users any given second. I imagine their spikes are significantly higher than their slow periods, too.

So a bit less than 6 live connections per CPU saved with the garbage collector configuration.

Well sure. And I've worked on software run by ISPs that have had more than 4M concurrent active sessions. Handled by something like 20 machines. Yes, they probably have way more overhead per session, but it is still a really high amount of servers.

Reply on news.ycombinator.com