PolarDB-X AI node storage options and recommendations-PolarDB(PolarDB)-阿里云帮助中心

This topic describes the storage options supported by PolarDB-X AI nodes, helping you choose the appropriate storage solution for different requirements such as model files, shared data, persistent data, and temporary cache. Pods on AI nodes can mount storage through Kubernetes PersistentVolume (PV), PersistentVolumeClaim (PVC), and emptyDir. Storage resources must match the network environment of the AI node, including the Kubernetes cluster, VPC, and region.

Supported storage options

Storage option	Use case	Access mode	Data lifecycle
ESSD cloud disk	Model caching, intermediate inference results, KV Cache, and data that requires persistent read/write	Typically suitable for a single Pod or a small number of Pods	Cloud disks exist independently of Pods and are not released by default after the Pod is deleted.
CPFS	Large model weights, high-throughput shared files, and shared read/write across multiple Pods	Multiple Pods can share access	Data is stored in the CPFS file system and is not deleted when Pods are removed.
OSS Bucket	Model weights, configuration files, shared read-only data, and archived data	Multiple Pods can share access	Data is stored in OSS and is not deleted when Pods are removed.
NAS	General shared files, configuration files, and lightweight shared read/write data	Multiple Pods can share access	Data is stored in the NAS file system and is not deleted when Pods are removed.
Local ephemeral storage	Temporary compute files and scratch data	In-Pod access	Follows the Pod lifecycle and is released when the Pod is deleted.
Memory-backed temporary directory	`/dev/shm`, high-frequency temporary read/write, and small file caching	In-Pod access	Uses available Pod memory and is released when the Pod is deleted.

Selection recommendations

Requirement	Recommended solution
High-performance persistent block storage required	ESSD cloud disk
Multiple inference Pods sharing large model weights with high throughput requirements	CPFS
Multiple Pods sharing model files, configuration files, or archived files	OSS Bucket
General shared file system capabilities required	NAS
Only temporary compute files need to be stored	Local ephemeral storage
Need to mount `/dev/shm` or use a memory-backed temporary directory	Memory-backed temporary directory

Limits

Storage resources must be network-connected to the cluster where the AI node resides. We recommend that you select storage resources in the same region and VPC or accessible over the internal network.
ESSD cloud disks, CPFS, OSS, and NAS all require creating a PVC first and then referencing the PVC in the Pod.
The performance and consistency semantics of shared storage such as OSS, NAS, and CPFS are determined by the corresponding cloud product. Select based on your business read/write model.
Local ephemeral storage and memory-backed temporary directories do not provide persistence. Do not store model weights, business results, or other data that requires long-term retention.

Use ephemeral storage

If you only need to store temporary compute files, you can use Kubernetes emptyDir.

apiVersion: v1
kind: Pod
metadata:
  name: ai-node-scratch-example
  namespace: default
spec:
  volumes:
    - name: scratch-volume
      emptyDir: {}
  containers:
    - name: app
      image: <image>
      volumeMounts:
        - mountPath: /scratch
          name: scratch-volume

If your application requires /dev/shm, you can mount a memory-backed temporary directory by setting emptyDir.medium: Memory.

apiVersion: v1
kind: Pod
metadata:
  name: ai-node-shm-example
  namespace: default
spec:
  volumes:
    - name: shm-volume
      emptyDir:
        medium: Memory
        sizeLimit: 32Gi
  containers:
    - name: app
      image: <image>
      volumeMounts:
        - mountPath: /dev/shm
          name: shm-volume

Set sizeLimit based on application concurrency and memory consumption during model execution to avoid excessive memory usage that affects Pod stability.