Configure Fluid Dataset Access Modes for ML Training Pipelines - ACK

In the Fluid framework, the default access mode of a dataset is ReadOnlyMany. You can modify the accessModes field in the Dataset spec to change the access mode. Fluid supports two access modes, ReadOnlyMany and ReadWriteMany. Set the accessModes field to match your workload's data requirements.

Access modes

Mode	Behavior
`ReadOnlyMany` (default)	Pods on nodes where the dataset is mounted can only read data.
`ReadWriteMany`	Pods on nodes where the dataset is mounted can read and write data.

Use cases

ReadOnlyMany — Use when your workload only reads data, such as training machine learning models or loading model files from Object Storage Service (OSS) or File Storage NAS (NAS) for model inference.
ReadWriteMany — Use when your workload reads and writes data, such as reprocessing input data and writing the results back to a caching system or backend file system.

Set the access mode

Set accessModes in the Dataset spec before you create the dataset. The following example configures a Dataset with ReadWriteMany using JindoRuntime (JindoFS) as the runtime.

apiVersion: data.fluid.io/v1alpha1
kind: Dataset
metadata:
  name: demo-readwrite
spec:
  mounts:
    - mountPoint: <mountpoint>  # Format: oss://<oss_bucket>/<bucket_dir>
      name: demo
  accessModes:
    - ReadWriteMany  # Valid values: ReadOnlyMany (default), ReadWriteMany

Parameter	Description
`mountPoint`	The data source mount point. For JindoRuntime, use the format `oss://<oss_bucket>/<bucket_dir>`, where `<oss_bucket>` is the name of your OSS bucket and `<bucket_dir>` is the subdirectory (defaults to the root directory).
`accessModes`	The access mode for the dataset. Valid values: `ReadOnlyMany` (default) and `ReadWriteMany`.