How useful this is in practice isn't something I've tried to measure myself, but given the performance variance people report it might not be a bad idea to try for yourself.
How useful this is in practice isn't something I've tried to measure myself, but given the performance variance people report it might not be a bad idea to try for yourself.
The article also doesn't mention spot pricing and per-second billing. Being able to burst up to hundreds of machines in minutes is super nice. I did some back of the envelope calculations for some build servers and it came out favorably for $cloud providers. Having idle machines also cost money.
For a highly optimized and predictable work load I'm sure providers like OVH can provide better bang for bucks. But that also assumes you don't use any of the cost reduction methods that are available with $cloud, like using spot pricing for variable load and reserved capacity for base load.
It's also work you simply don't have to do if you rent hardware.
I'll be frank: If you haven't seen performance variance with physical hardware, you either have been incredibly lucky, not paying attention, or not working at a very large scale.
Also neighbours will change over time so testing for 24h and committing to a volume long-term doesn't sound optional.