Ask HN: Opinions on serverless compute for ML training
I'm evaluating serverless compute options for ML training, most notably Azure's serverless compute (since I have free Azure credits): https://learn.microsoft.com/en-us/azure/machine-learning/how-to-use-serverless-compute?view=azureml-api-2
So far, I've always managed my own instances. Upside of that approach is no hidden abstraction layers and excellent pricing as I can use spot instances and pricing on VMs in general is very competitive and easy to understand. Downside is that it's somewhat inflexible/manual when it comes to firing up different instance types or groups of instances, and I often forget to shut down the instances which ofc is quite expensive.
I'm having a hard time understanding which of the offerings are worth using, and which ones I'd be wasting my time with. What do you recommend?