>>> how unremarkable their architecture is – their level of traffic is comfortably running on 4 MS SQL Server boxes:

Wrong, and wrong.

First, their architecture is extremely remarkable. They are doing and mastering vertical scaling, down to every little details. Terabytes of RAM, C# instead of Ruby/Python, FusionIO drives, MS SQL instead of MySQL/PostGre, etc...

Second, that's at least 6 database servers, each one being more expensive than 10 usual commodity servers:

First cluster: Dell R720xd servers, 384GB of RAM, 4TB of PCIe SSD space, and 2x 12 cores. It hosts the Stack Overflow, Sites (bad name, I’ll explain later), PRIZM, and Mobile databases.

Second cluster: Dell R730xd servers, 768GB of RAM, 6TB of PCIe SSD space, 2x 8 cores. This cluster runs everything else. That list includes Careers, Open ID, Chat, our Exception log, and every other Q&A site (e.g. Super User, Server Fault, etc.).

I think you misunderstand my use of unremarkable. It's not wrong or bad but key parts are the same stack you'd have picked 1-2 decades ago – well implemented, values scaled up, yes, but something you could have shown people in 2000 and not had to explain more than that the hardware costs so much less.

I think that's good: SQL databases are very mature and you don't want to be exciting for your core business data if you don't get some major benefit to defray the cost. Boring is a delightful characteristic for data storage.

My bad then. I think that using a proven technology perfectly tuned and the way it's meant to be done, is remarkable. This should be promoted more, instead of the new and shiny.

I'm not sure you could get 1TB of memory and multi TB SSD drives anywhere in the 2000's, even for a million dollar. That makes a major difference in the ability to scale up. Data didn't grow, storing 1M user account always took the same space.

Agreed that data sizes would have been far more of a challenge – far more people used clustered services just because a single box only supported so much RAM, RAID arrays for IOPs and size as well as redundancy, etc. – so it's arguably become much easier for a growing chunk of the industry.

>believe they'd need a clustered NoSQL database or it wouldn't scale.

While for many developers, NoSQL might be overkill - Stackoverflow is a bad example. If you were any fast growing startup in the cloud, and you wanted to go the SO route it would have meant going CoLo. SO has set of machines with a nearly a terabyte of RAM - GCE doesn't even offer cloud machines with the same specs.

And even then their setup is far more fiscally expensive than something you could get done with 20 cloud nodes on some NoSQL solution.

One: run the numbers on the dozens of cloud nodes you mentioned. Is that really cheaper than renting some servers in a colo?

Second: how much time would they have spent rolling all of the data integrity, reporting, etc. features they'd have needed to add. I'm inclined to take them at their word when they say this was safer and cheaper given their resources.

>I'm inclined to take them at their word when they say this was safer and cheaper given their resources.

I never meant to claim that cloud was safer and cheaper. What I meant was, for the majority of operations, staying in the cloud with some distributed setup is likely more feasible that moving to CoLo (see GitLab).