Skip to content

Storage

VAST Data Platform

PAIS uses a VAST network-attached storage system as the shared storage backend. It provides high-throughput, low-latency access to large datasets, model checkpoints, and experiment artifacts across all nodes simultaneously.

Property Value
Technology VAST Data Platform
VIP 10.60.200.240 (C302-pool)
Management endpoint 10.64.36.10
Total allocation 10 TB (model cache) + group quotas
Access method VAST CSI Driver 2.6.5 via Kubernetes PVCs
Reclaim policy Retain (data survives PVC deletion)

Storage classes

StorageClass Reclaim policy Use
vast-retain Retain Research data, model weights
vast-retain-rwo Retain ReadWriteOnce workloads
vast-retain-rwx Retain ReadWriteMany (shared notebooks)

PVC layout

PVC Namespace Mount path Purpose
model-cache-pvc pais-inference /model-cache HuggingFace model downloads
rg-compsci-data rg-compsci /vast/rg-compsci Group datasets and results
rg-abi-data rg-abi /vast/rg-abi Group datasets and results
mlflow-artifacts mlflow /mlflow/artifacts MLflow run artifacts

Root-squash behaviour

VAST uses root-squash: processes running as root inside containers have their UID mapped to nobody (UID 65534) at the storage level. This can cause unexpected ownership on files written by privileged containers.

Mitigation: The pvcviewer-uid-patcher deployment ensures PVCViewer pods run as UID 1000. For other cases, apply appropriate file permissions after writing:

chmod -R 755 /vast/rg-compsci/my_output_dir