Databases and copies
On this page
Database statements are sent to any server and act on the whole cluster.
Creating and dropping
CREATE DATABASE orders;
CREATE DATABASE ledger OPTIONS {replicas: 5};
DROP DATABASE orders;
- A new database is held by three servers, each with a vote, chosen from the servers that hold the fewest databases.
- Every server serves every database, whether it holds a copy or not.
- A dropped database is deleted from every server that held it. One created later with the same name is a new, empty database.
- The default database is kept; it is not dropped.
How many copies
replicas is the number of copies that vote. It is odd: three survive the loss of one
server, five the loss of two. An even number would cost a server and survive no more.
ALTER DATABASE orders SET OPTIONS {replicas: 5};
Copies are added or taken away to match while the database serves. When the cluster has fewer servers than a database asks for, the database runs on those there are and is made up as servers are added.
Where the copies are
SHOW PLACEMENT;
SHOW PLACEMENT FOR DATABASE orders;
CALL gdb.cluster.members('orders');
| Shows | |
|---|---|
SHOW PLACEMENT | Which servers hold each database, and whether each copy votes |
gdb.cluster.members(database) | Each member: node, address, role (leader, follower or learner), voter, and how far it has got (position, behind) |
Moving a copy
The cluster evens its servers out by itself. To move one copy yourself:
CALL gdb.cluster.move('orders', 2, 4);
The copy is made on server 4 first and only then removed from server 2. If server 2 leads the database it hands leadership on before its copy goes. The database serves throughout.
Placing copies by hand
These procedures make one change to one database's members. They are what the cluster
itself uses. Pause it first with CALL gdb.cluster.pause(), so that it does not put back
what it would have chosen, and CALL gdb.cluster.resume() afterwards.
| Procedure | Does |
|---|---|
gdb.cluster.addMember(database, node) | Gives a server a copy, without a vote |
gdb.cluster.promoteMember(database, node) | Gives a member a vote once it is within 1,000 transactions of the leader |
gdb.cluster.removeMember(database, node) | Takes a server out of the database's group. The leader hands leadership on first |
gdb.cluster.transferLeader(database, node) | Hands leadership to a member that votes |
A change that is refused fails with Galactus.ClientError.Cluster.ChangeRefused and says
why.