Ah ok ! I think the link between clearml agent queues and the autoscaler is also important for resource monitoring etc. Sorry I'm quite new to clearml so I'm still trying to understand the architecture 🙈 . Although AWS Batch has its own queue system too which it uses to allocate jobs on training instances but its not customizable

This is a nice issue to open in https://github.com/allegroai/trains :)

Hi WackyRabbit7

Currently we don’t have a GCP auto scalar. We’re more than happy to get contributions for GCP and other platforms.

The AWS auto-scaler code is pretty generic and in order to be used for GCP, you need to implement a  GCPAutoScaler  similar to  trains.automation.aws_auto_scaler.AWSAutoScaler  , which basically has  spin_up_worker()  and  spin_down_worker()  methods...

Hi  TimelyPenguin76 ! quick question - I was curious if at any point in time ClearML had considered using AWS Batch for the autoscaling part ? As submitting "training jobs" to AWS batch would create/terminate ec2 instances automatically too instead of clearml writing logic to spin up/down instances

Hi BrightElephant64 , can you add an example? Also, the ClearML AWS autoscaler know how to work with ClearML-agent queues

I'd be lying if I said I had time for that 🙂