Oh I am considering to us FSx for Lustre with S3 export/import for our ML workloads. Isn't it stable enough?