# Nebula Cluster

[Wiki](../README.md)

Nebula Cluster runs your databases on several servers at once. Each database is kept as
copies on three or more of them, a write is acknowledged once most of those copies hold it
on disk, and when a server is lost another copy takes over without anyone stepping in.
Applications connect to any server over plain Bolt and use the cluster as they would one
server.

Nebula Cluster is in **early access**, for Enterprise. These pages describe Nebula
3.0.0-alpha.1. [Talk to us](https://galactusdb.com/contact) to run it.

## Articles

- [Start a cluster](start.md): what each server needs, the settings, and the first start.
- [Connect applications](connect.md): the addresses an application can use, the router,
  what a session sees and the codes to retry on.
- [Databases and copies](databases.md): creating a database, how many copies it keeps,
  where they are and how to move them.
- [Servers](servers.md): adding a server, taking one out, what happens when one is lost,
  and how the cluster keeps itself in order.
- [Back up and restore](backups.md): full and incremental backups taken while the database
  is in use, restoring to a moment, and checking a backup.
- [Monitoring](monitoring.md): metrics for each database's copy, the readiness check and
  the list of servers.
- [Property shards](property-shards.md): spreading a database's properties over groups of
  servers, for a database larger than one server's disk.

## How it works

| Part | How it works |
|---|---|
| **Copies** | Every database is held by an odd number of servers, three by default. One copy leads and takes the writes |
| **A write** | Is acknowledged once more than half the copies have it on disk. Three copies survive the loss of one server, five the loss of two |
| **A read** | Sees every write acknowledged before it, on whichever server it runs |
| **A lost server** | Another copy takes over the databases it led. Its copies are made again on other servers after ten minutes, which is a setting |
| **Any server** | Takes any request. One that does not hold the database, or does not lead it, carries the request to the server that does and returns the answer on the same connection |
| **The control plane** | The cluster places copies, repairs them and evens them out by itself. Whichever server leads the cluster's own records does this; there is nothing else to install |
| **The router** | A small program that gives applications one address. It is a choice, since every server takes any request |

Under the plain statements: each database's copies are a Raft group, the group's log is the
database's own commit log, and a server proves who it is to the others with a certificate
from an authority you run.

## Related articles

[Replica databases](../operations/replicas.md) · [Server configuration](../operations/configuration.md) · [Backup, restore and transfer](../operations/backups.md)
