Nate Atkins
02/11/2020, 4:28 PMJoe Schmid
02/11/2020, 4:38 PM"dask-resource:KEY=NUM"
(See: https://docs.prefect.io/api/latest/engine/executors.html#daskexecutor) We use this successfully for both GPUs and for what we call "High Memory" Dask workers, i.e. to execute Prefect tasks that need a lot of RAM. Note that to make this work from the Dask side will depend a lot on how you run a Dask cluster that Prefect executes on. We use a long-running Dask cluster on AWS EKS with different node groups for 3 types of Dask workers: "regular", GPU, and "high memory" and then scale those via Deployment replicas.Nate Atkins
02/11/2020, 4:40 PMJoe Schmid
02/11/2020, 4:43 PM