Skip to content
/ sion Public

Distributed file system for many large files

Notifications You must be signed in to change notification settings

RaasAhsan/sion

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

sion

Sion is a scalable, distributed file system designed to reliably store and serve millions of files to thousands of clients. It is designed to support files that are terabytes in size by distributing the workload over many machines with heterogeneous hardware allocations. The storage capacity and aggregate bandwidth of the cluster scales with the number and size of machines and disks that participate in the system.

Unlike most distributed file systems, Sion is append-only. This design principle considerably simplifies the architecture of the system, while still supporting many data workloads with high reliability. In particular, readers and writers never block eachother during any operation while preserving proper isolation and integrity.

Sion does not expose a traditional POSIX file system API to clients. Instead, clients interact with it via several HTTP+JSON APIs through which all control plane and data plane operations are performed.

Architecture

A Sion cluster contains a primary master which manages the state of the cluster (the control plane), and many storage nodes which store and serve files (the data plane).

Master

A Sion master performs three major processes: cluster management, namespace management, and placement. Each process is compartmentalized into its own subsystem:

  1. The cluster subsystem processes cluster operations and tracks the state of the cluster.
  2. The namespace subsystem processes file system operations and serves this information to storage nodes and clients.
  3. The placement subsystem makes assignment decisions for where chunks should be located. Placement decisions are used to respond to failures in the cluster and balance load across nodes.

Subsystems often need to interact with eachother in response to certain requests or events. For example, when a node dies, the cluster subsystem will instruct the placement subsystem to reschedule all its chunks. This communication primarily takes place over a Go channel.

Notes

$ curl -XPOST http://localhost:8000/files -H "content-type: application/json" --data '{"path":"path"}'

About

Distributed file system for many large files

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published