What’s the best way to estimate cloud costs (GCP) for an MLOps system in production?
Here is another installment where we take some of the best threads from our slack channel and make them available to anyone. If you want to join our slack channel you can do so here.
What’s the best way to estimate cloud costs (GCP) for an MLOps system in production?
Hi. If it were AWS, there’s a good pricing calculator you could use… GCP I’m not very familiar with. Sorry to not help…
All major clouds have calculators; I would suggest running a small prototype, perhaps at 3 different usage levels, get the resource usage report (not the cost as you are likely to be in the always-free tier) and extrapolate (in sheets/excel or causal) then plug that back into the calc and see what the cost will be.
@Dmitriy its going to depend on your system tbh, I can provide you with a high level cost estimator for a DIY k8s based platform. If your using the ML as a service from the different clouds they are pretty obfuscated (ie sagemaker charges on 5 different dimentions).
but super high level a MLOps systems with decent traffic, high availability, back up recovery, metrics sevices would be $3-$4k in total cloud cost a month
you can chip away at some of that with no GPUs etc
calculator.aws
AWS Pricing Calculator lets you explore AWS services, and create an estimate for the cost of your use cases on AWS.
You can estimate cloud costs using the GCP calculator but I believe it all depends on the tools and resources you intend to use (e.g. GKE, GCE, GCS, highmem machines, GPU, etc).
Google Cloud Platform Pricing Calculator
Create your own Custom Price Quote for the products offered through Google Cloud Platform based on number, usage, and power of servers (17 kB)
Great convo so far! It really is a case by case situation. You have to constantly assess costs that can creep up such as GPU consumption or data transfer fees.
Agree with @Diego Oppenheimer if its DIY Kubernetes vs. managed services prices can be highly different prices. If it is Kubernetes you are talking about, worth checking out this calculator: https://www.getoctane.io/calculator and you can pass in k8s yaml to estimate cloud cost (edited)
Love this thread! Really appreciate all the inputs!
@Demetrios I believe this thread deserves a post on Medium.
Check our weekly MLOps meetups, or listen retroactively on youtube or podcast land.