undefined | Better HN

0 pointsrahimnathwani9y ago0 comments

Instead of 2 boxes that can handle all your load, why not 4 that can each handle 1/3rd of the load?

0 comments

4 comments · 2 top-level

user59944619y ago· 2 in thread

Let's say that one box is down (don't care why, boxes die all the time).

You have 3 boxes, they're all running at 100%, or worse if the load is not perfectly balanced (it never is). You've got bad latency and you're up for cascading failures. Your system is in bad shape.

Now, supposing you have to deploy your application. Gotta take a server offline temporarily, the load will be 100% on the last 2 servers left... that can only take 66%. Booya! You're down :p

That's what happen when people are cheap on servers, in high performance systems.

vidarh9y ago

Yes, but his point is still valid: You can to an extent reduce the overprovisioning you need by adding more, smaller units.

In your example, two boxes would have meant downtime no matter the capacity of each box too. All your example demonstrates is a reason why you might want to have capacity that allows taking more than one box offline....

Which could be handled with bigger servers, or more smaller servers.

user59944619y ago

Yes to all of this.

I write this to illustrate that you must have a significant batch of servers (about 5) in order to achieve costs savings (by using smaller servers) WHILE not affecting reliability.

Well, that is, if you care about rare cascading failures and performances issues.

aaronblohowiak9y ago

yep, and it is cheaper as well.

j / k navigate · click thread line to collapse