> If you have 1 massive job, Ray sucks and doesn't provide table stakes.

Can you say more?

I'm really glad to see a post like this come out. I've seen so many discussions online about customizing models -- this post really does cut through the noise.

Really like the evaluation methodology, and seems well-written as well.

This is a huge release! Weights and Biases support, transformers, Ray support -- it just never stops!

Interesting. What do you mean when you say: "versus the authors re-applying some of same research in their own thing."

And does "code distribution" just mean to release as open source?

code distribution I mean like cloudpickle. Which Ray has, but I wonder about an equivalent to Spark files (torrented and cached). A Kubernetes runner is also important ...

I wish the Ray peeps would consider just trying to merge some with the Spark RDD API. Reynold (at Databricks) is kinda hard to deal with, but so far to me it looks like the aim of having Ray team build things from the ground up has simply re-validated a lot of the systems work (but not all) that’s already in Spark.

Very related open source project from the RISE Lab at UC Berkeley: http://ray.readthedocs.io/en/latest/tune.html

Very cool project! Glad to see support for PBT and HyperBand. Also related to https://www.microsoft.com/en-us/research/publication/hyperdr...