Hacker News Reader

Comment by jackowayed | Hacker News Reader

For now they really only saved $300k/yr, which is less than the cost of two engineers. Though I guess cutting a nontrivial expense that is likely to grow by nearly a factor of three is pretty great and probably worth investing in.

That said, this adds complexity to their systems with the only benefit being cost savings. Given that we can assume that no code is perfect, it's likely that at some point the auto-downscaling will cause an outage or period of slow responses, which could easily lead to lost usage and trust that costs them as much as they're saving on ops.

View on HN

IMO designing systems that can be powered on and off regularly (and quickly) is a good thing in itself; it encourages "proper" setup. I've found servers that have been on for months or years tend to need manual intervention after a reboot. When the machine could reboot every day, you can't have that.

In other words, it's just good design.

Nitpicking: but it's not good "in itself", it just has other benefits.

> IMO designing systems that can be powered on and off regularly (and quickly) is a good thing in itself

Despite the fact that, in theory, a mainframe should never go down, most dinosaur pens will power cycle them regularly just to see what happens if you come up from a cold start.

A mate of mine worked in a dino pen where they did this on Saturday evenings. He told amusing stories.

> That said, this adds complexity to their systems with the only benefit being cost savings.

That's not the only benefit. They also get much better reliability. The systems scale down with load, but they also scale up. As their load increases their system scales up along with it, giving greater reliability during increased load. Since they have to architect for that, it makes scaling up in general easier.

> For now they really only saved $300k/yr, which is less than the cost of two engineers.

It took them 2 weeks to implement. That means it's saved a lot more than it took to implement (2 weeks times a couple of engineers) already.

To use your metric, if they saved close to $300K/year, that means they could afford to add 2 engineers to their staff, which is very significant.

Is cost savings really the only benefit? It seems to me that this is a better engineering outcome.

That depends on for who. For Amazon, perhaps.

Better for Pinterest too. It means they need to develop services that can be launched and terminated quickly, and that their services are much more resilient during times of high load.

Reply on news.ycombinator.com