Some time ago, I evaluated the pricing for someone who was thinking of moving a web server from a traditional private server to AWS. It seemed that the monthly cost would be lower, until I looked at the bandwidth cost. The monthly price of that traditional private server, which included a fixed bandwidth limit, ended up being lower than what it would cost just for the bandwidth on AWS.
One killer no one seems to notice is the bandwidth between availability zones. When you have a proper best practice cross AZ deployment it can be rather expensive. But still cheaper than switches and humans and cages and data centre contracts.
https://www.computerworld.com/article/3044261/dropbox-quits-...
Due in part to an extremely large number of hidden factors.
I'm not just talking about direct cost, but even 1vCPU != 1vCPU.
When discussing s3/gcs storage; for instance, you would need to understand the difference between each of the storage classes and how they're charged for access.
This would look much more like a series of a million graphs than a table.
On the whole: I'm extremely unimpressed by this article. Maybe one day when I'm bored enough I'll give a shot to a reasonable cost comparison of some actual infrastructure.
Maybe even throw in some TCO calculation for administrative duties, such as understanding your billing, ease of debugging and setup. If I really hate myself that day.
And the vCPUs change at providers too. It used to be that AWS Lightsail was consistently the poorest performing vCPU in our tests, but yesterday their 4vCPU instances were beating other cloud providers. Despite ostensibly having a slower clock speed according to /proc/cpuinfo
For other services, not so much. Like, is 1 CPU / Memory in Fargate the same as in App Runner and Lambda?
Hell, Brendan Gregg, father of observability, noticed a 10x performance difference of machines of the same type inside AWS and that’s before talking about different providers.
GCP is also quite transparent with their CPU capabilities. But I would really hesitate to say 1x2.2GHz vCPU in GCP is equal to 1x2.2GHz vCPU in AWS
For example our S3 costs are higher than our EC2 costs and we made a productive saving implementing lifecycle and S3 IA migration. The biggest EC2 cost reduction we made was through careful analysis and moving some stuff to Lambda, not by changing provider or instance class optimisation.
Whatever happens though, you don't know what the costs are going to be until you get the first bill, regardless of what you estimate.
For example, we rely heavily on storage (with deployments in the 100TB - 1PB range), where we have a lot of churn in data and as such require a lot of throughput. AWS’s GP3 EBS volumes offer 1GB/sec throughout each at a really attractive price, there simply isn’t any comparable offering for this at Azure at that price point (only the ultra premium SSD variant, which is more like IO2 EBS).
Does anyone know whether there are any real in-depth studies and comparisons between performance of cloud providers, on the level of, say, the STAC benchmarks?
This annoyed me, so I closed the article. How's that for engagement?
IMO:
- cloud is expensive with large resource needs, so bare metal is better
- cloud is hard to use for few deployment needs, so bare metal is better
So, the best case to use cloud is when you require lots of deployments but relatively small resource needs for each. For example, testing a piece of software on many platforms.
But if you do a 1 year commitment it's $0.10...
I think they got the data messed up in here.
Short answer: it depends
Happy to get any feedback on this!
