top | item 24590499

Yesterday's Google outage was a pool of routing servers crashing

124 points| tyingq | 5 years ago |twitter.com | reply

32 comments

order
[+] Kaknut|5 years ago|reply
Dependencies of these big providers like Google, Microsoft, Cloudflare are increasing which results to failure on a wide scale even if one fails. Distribution is the key.
[+] auganov|5 years ago|reply
Well for the vast majority of simple apps you're better off failing when everybody else is. People will blame it on you less. When your alternative solution fails and everything else seems to be up the blame will fall on you.
[+] tyingq|5 years ago|reply
Google could probably do a better job here and not put so many services on the same pool of L7 devices. Separate pools with smaller groupings would reduce the blast radius.
[+] heartbeats|5 years ago|reply
Anyone know what routing servers? BGP?
[+] kyrra|5 years ago|reply
(Googler, opinion is my own, I know nothing about this specific outage).

Google has LOTS of internal routing systems. BGP is about anouncing what IPs a given network can handle, which is not the case here.

Before hitting application level routing, I believe you hit the Maglev[0]. Seems unlikely this was the cause, as it would likely take down all services.

One of the first application layers balancers you hit that is well known is the GFE[1][2]. This is similar to an HTTP reverse proxy, but Google made. I could definitely see this as the cause.

[0] https://static.googleusercontent.com/media/research.google.c...

[1] https://cloud.google.com/security/infrastructure/design#goog...

[2] https://landing.google.com/sre/workbook/chapters/managing-lo...

[+] MyelinatedT|5 years ago|reply
Sounds more like an application load balancer issue ("routing requests" seems to imply L7) than network routing, but I might be misunderstanding.
[+] enneff|5 years ago|reply
Traffic entering Google's network hits a bunch of front ends that route traffic to the relevenat back ends. I'd guess it's those application-level front ends that were having trouble, rather than anything network-level like BGP.
[+] skim_milk|5 years ago|reply
There's a huge """secret""" Google data center in Council Bluffs, Iowa that appears to be in the finishing phases of completion. I talked yesterday to a union worker who is moving to Des Moines to work on a new Microsoft data center there tonight, it appears that work is drying up at this data center here and a lot of the travelling blue collar folk are leaving this area.

I wonder if this data center coming apparently partially online is a part of the problem?

Also, after this he is likely to work on an Amazon fulfillment center next year - impressed by all the (albeit temporary) blue collar jobs created by FAANG at the moment!

[+] SteveNuts|5 years ago|reply
I think by now Google has lots of experience with bringing data centers online, so this seems incredibly far fetched to me.
[+] qmarchi|5 years ago|reply
Not super well known, but Google has quite a few Datacenters that aren't used for Google Cloud, but are reserved for internal use.
[+] rezonant|5 years ago|reply
Oof. This DC has been around since 2007. I can make a VM in it on GCP right now :-)