File & Object Storage
MASS distributes application data across NVMe storage nodes. Applications access that data through a POSIX filesystem or an object API, depending on their integration.
Volume Groups and Volumes
A volume group provides shared capacity for its volumes. Volumes organize application data and define access permissions and protection settings. This lets administrators manage datasets separately while drawing on the same storage capacity.
Distributed Data Access
The client determines where to send an operation using data placement information and the cluster's topology. It then communicates with the storage nodes responsible for the data. The Dashboard is not involved in each read or write.
Data is distributed across storage nodes so multiple clients can access different portions in parallel. Operations that reach the same device, network link, or shared data can still compete for resources.
Data Protection
Replication stores multiple copies of data. Erasure coding stores data fragments and additional parity fragments that allow data to be reconstructed after a failure. Both consume extra capacity; the selected policy determines the balance between usable capacity and protection.
| Method | How it protects data | Capacity consideration |
|---|---|---|
| Replication | Keeps copies in separate failure domains | Two copies require roughly twice the data capacity |
| Erasure coding | Reconstructs missing fragments using surviving data and parity | Overhead depends on the data-to-parity ratio |
Protection depends on the policy, placement, and number of available failure domains. During recovery, rebuilding data also uses storage and network resources.
Relationship to KV Cache Storage
File and object volumes retain application data. The KV cache store holds reusable inference results and uses its own data path. An unavailable cache page can be recomputed by the inference engine; application data is retained according to its storage protection policy.
For the factors that affect throughput and latency, see Performance Characteristics.