Hey everyone 🙂 Can somebody explain to me how metaflow with aws_batch chooses the instance type in case multiple option match the requested ressources ?

That would depend on the strategy that you choose on the AWS Batch end

thanks ! Didnt want to match GPU instances by accident just because their cpu and ram matchted nicely 😄

You could create multiple compute environments and set the priority on the GPU queue the lowest

For a starter, here is a Cloudformation that deploys separate queues as Savin suggested.

At the moment, AWS Batch doesn’t provide a mechanism where you can set the CPU jobs running on GPU nodes as pre-emptible so that they are evicted as soon as a GPU workload shows up in the queue

@crooked-jordan-29960 just curios is there a reason why the version in the official cloud template is quite old ? ('netflixoss/metaflow_metadata_service:v2.3.5') i replaced it with a newer version, which seems to just work fine too, but i just started using metaflow 2 days ago 🙂

No interesting reason - tried to minimize edit distance from the standard metaflow deployment (no gpu) at that time. If a newer version is working well and you'd like to make a PR we can update it!

Alright, thanks for the infos ! 🙂