Inside a Google data center

What people said

ZeroSolstice · 2023-11 · Reference

50 SRE's sounds like they could keep borg running. There really isn't that much details on what borg is made up of outside of a 2015 research paper[1]. Just to address the parts you laid out. 1. Keep Borg running What would entail running? There are plenty of small teams that manage hundreds of servers and virtual machines through monitoring, deployment, networking, decom, etc lifecycle. This seems counter to purpose of cloud computing, commodity hardware and abstracted compute resources that can be deployed on-demand. Not to mention redundant sites, availability zones, <insert your cloud providers name for High Availability(HA) feature here>. Is this the part that SRE's maintain? 2. Machine maintenance From documentaries[2] and other comparable data center operations this would seem to be handled by on-site datacenter staff. Disk replacement, physical replacement, network cabling, etc. Without any other operational info I would doubt resources would just be dedicated to borg as its a bit of a overall technician site work. As for OS and software updates and again without any other info or insight would seem to fairly automated after passing testing or at least passed to another team that just handles updates. 3. Machine procurement Again commodity hardware or if its all custom still that would seem to be more of a EE task for build out and accounting for purchasing. Otherwise at Google's scale you would just get pre-populated racks delivered and replace the entire rack when 5

Read the original HN comment ↗

ZeroSolstice · 2023-11 · Reference

I appreciate the examples but none of these items are difficult regardless of size of the organization. These are standard system and network administrative items or on-site technician work. Unless you have other data about the borg architecture to show, its entire premise from the paper[1] is that it just runs jobs and removes the developers needing to know about the underlying hardware or OS that its running on it. From the paper I referenced its huge cluster and if its their own software I'm sure with the development experience at google they have accounted for loosing a cluster node. Losing any number of cluster nodes is the same scenario if you were performing an software or hardware upgrade in which case you would be taking a node offline. I would be very skeptical that there are people manually kicking off upgrade jobs for nodes in the cluster. Maybe an A/B deployment or burn-in for a week before fully pushing it out with a pipeline workflow. I'd more so say that their borg implementation probably runs at different minor versions frequently. Google runs their own data centers, thats the video documentary I referenced[2]. For machine procurement I don't see that as a SRE task. If they are making custom boards and systems they are going to have an entire group supporting that. At the size of google I'm thinking they wouldn't be running into supplier issues that they could figure out easily or account for in their borg software. Again its supposed to be one big abstractio

Read the original HN comment ↗

jeffbee · 2021-12 · Joke / meme

I guess you thought wrong. What do you think this hardware tech is doing in this video? [VIDEO] Men with carts scale perfectly. One guy can manage 1000 machines. 2 guys can manage 2000. Perfect scaling.

Read the original HN comment ↗