[greenleafjacob]

At Uber we invested in synthetic load tools that create fake riders and drivers, match them, etc and test our entire dispatch system end to end to arbitrary amounts of online driver and riders. I don’t see why they couldn’t do the same with carts, adding products to carts, etc

[wiredone]

They do. But I'm guessing you don't scale up 1M+ servers for your Uber canary traffic tests - these are some of the scales Amazon undergoes in these events. The scale is unlike almost any other web property around.

[BinaryIdiot]

> But I'm guessing you don't scale up 1M+ servers

Do they really need 1 million servers? Many of my friends who work at other tech companies need such few servers in comparison even with significantly high traffic that just screams massive inefficiencies...which seems wrong.

But I've never worked at Amazon so I wouldn't know.

[convivialdingo]

At one of my previous companies we managed about 100 servers for peaks of about 30 million users.

But we didn’t handle free-form text search like Amazon. I can imagine that would necessitate a huge scaling up of compute and data.

[deegles]

Judging by all the error pages after launch, they needed more!

[greenleafjacob]

I’m not sure how it’s relevant. If you have the infrastructure to send 1000 concurrent users you can probably send 1M concurrent users. We only test small integer multiples of our peak traffic, and if your absolute number of servers to service that is in the millions then it would make absolute sense to be routinely running that capacity test. If that means “scale up 1M+ servers” then that is what you have to do, otherwise how can you be sure?