This document explains some basic assumptions and design approach, limits of the disk caching feature. If you're looking to get started with disk cache, we suggest you go through the [getting started document](https://github.com/minio/minio/blob/master/docs/disk-caching/README.md) first.
- The cache drives are required to be a filesystem mount point with [`atime`](http://kerolasa.github.io/filetimes.html) support to be enabled on the drive. Alternatively writable directories with atime support can be specified in MINIO_CACHE_DRIVES
- Expiration of each cached entry takes user provided expiry as a hint, and defaults to 90 days if not provided.
- Garbage collection sweep of the expired cache entries happens whenever cache usage is > 80% of drive capacity, GC continues until sufficient disk space is reclaimed.
- Cache-Control and Expires headers can be used to control how long objects stay in the cache. ETag of cached objects are not validated with backend until expiry time as per the Cache-Control or Expires header is met.
- To ensure security guarantees, encrypted objects are normally not cached. However, if you wish to encrypt cached content on disk, you can set MINIO_CACHE_ENCRYPTION_MASTER_KEY environment variable to set a cache KMS
master key to automatically encrypt all cached content.
Note that cache KMS master key is not recommended for use in production deployments. If the MinIO server/gateway machine is ever compromised, the cache KMS master key must also be treated as compromised.
Support for external KMS to manage cache KMS keys is on the roadmap,and would be ideal for production use cases.
> NOTE: Expiration happens automatically based on the configured interval as explained above, frequently accessed objects stay alive in cache for a significantly longer time.
Upon restart of minio gateway after a running minio process is killed or crashes, disk caching resumes automatically. The garbage collection cycle resumes and any previously cached entries are served from cache.
- Bucket policies are not cached, so anonymous operations are not supported when backend is offline.
- Objects are distributed using deterministic hashing among the list of configured cache drives. If one or more drives go offline, or cache drive configuration is altered in any way, performance may degrade to a linear lookup time depending on the number of disks in cache.