4 ms·
Google Cloud gets a +1 for this. When a big corp does the right thing for its customers, they take notice.
by ken47 10y ago
Google Cloud gets a +1 for this. When a big corp does the right thing for its customers, they take notice.
- s3v3n 10y agoAWS has had this for a while.
- kyrra 10y agoAWS I think has something similar, but I'm not sure it's the same, CloudWatch[0]. Cloudwatch giving alerting and can kill underutilized instances. But this sizing feature from GCP seems to allow for instance resizing rather than just deletion. As far as why this may benefit google: while a customer will pay the same regardless, that means there are less resources in GCPs overall pool of potential machines that others can use. And to support the overall capacity of all their customers, Google needs more machines to handle them all. I'm guessing Google would rather have most of their machines active, as building out less racks/datacenters is probably better for the bottom-line overall. [0] https://aws.amazon.com/cloudwatch/details/ https://aws.amazon.com/cloudwatch/details/
- vasco 10y ago"Trusted Advisor" on the AWS Console has been doing this for a long time, for several types of under-utilized, not only instances. Not only this but it also notifies you when you're reaching default limits so you can put in a support request for limit increase before you hit them.
- user5994461 10y agoYou gotta pay 10% on top of your bill to have Trusted Advisor. (that's part of the pro support plan). Last I checked there was no advice on instance size. I've seen one or two, nothing meaningful. There are lots of advice for reserving instances though, and they are extremely poor. Listening to them would be a good way for us to waste $10k/month :D
- sciurus 10y agoThe checks and recommendations aren't very sophisticated, though. E.G. the Low Utilization Amazon EC2 Instances check: Checks the Amazon Elastic Compute Cloud (Amazon EC2) instances that were running at any time during the last 14 days and alerts you if the daily CPU utilization was 10% or less and network I/O was 5 MB or less on 4 or more days.