Skip to main content
Dataset cache is an optional capability for accelerating data reads. It is useful when development environments or training jobs repeatedly read the same dataset.
If dataset caching is enabled for your project and the same dataset needs to be read repeatedly, you can create a cache.

​
How cache relates to datasets

  • Cache accelerates reads: Cache can improve access efficiency when the same dataset is read repeatedly.
  • Cache depends on source data: The source files in a Volume or object storage remain authoritative; cache does not replace source data management.
  • Cache consumes resources: The cache service continuously consumes CPU, memory, and other resources while it runs. Immediate preloading may temporarily consume additional instance resources.
  • Cache is not a backup: Clearing a cache does not delete the source data, but the cache itself must not be the only persistent copy.

​
Create a dataset cache

Before creating a cache, review the resources required to run the cache service and preload data. Plan the amount of data to cache based on the available capacity displayed on the page.
  1. Open the dataset cache creation page.
  2. Select the dataset to accelerate, then configure the resource pool, instance type, and quantity for the cache.
  3. Check the available cache capacity and decide whether to preload data immediately.
  4. After submitting, track the cache status and data volume, then verify the result through actual reads.

​
View cache status and capacity

If the cache percentage is below 100%, first check whether preloading is enabled, then compare the source data size with the cache capacity. With on-demand reads, data that has not been accessed may not yet be cached.

​
Manage dataset cache

Clearing a cache and managing source data are separate operations. Clearing a cache affects subsequent read performance, but does not delete source files in Volumes or object storage. The page not listing any users does not prove that no workloads depend on the dataset or cache. Before clearing or deleting a cache, review dependencies across training jobs, development environments, and other workloads.

​
Frequently asked questions

Check the CPU quota, resource pool availability, and the temporary instance resources that may be required for immediate preloading. If the page shows a failure reason, troubleshoot that reason first.
First check whether preloading is enabled, then compare the source data size with the cache capacity. With on-demand reads, data that has not been accessed may not yet be cached, so a cache percentage below 100% does not necessarily mean that the task failed.
Based on the status reason, check source data permissions, network access, and cache resources. Do not determine whether a failure occurred based only on the cache percentage.