Study Notes

CTDB, Cluster Trival Database.
1. Consistent data and consistant locking
2. If one node failed, CTDB will automatically recovery and repair all CTDB database.
3. CTDB is PCIFS, parallel CIFS with samba 3/4
4. Support TCP and Infiniband

Requirement:
This filesystem must be mounted and available on all nodes in the CTDB cluster.

HA Features:
The CTDB nodes in a cluster designates one nodes as recovery master through an election process. If the recovery master node failes a new election is initiated so that the cluster will always guarantee there will be a recovery master. The recovery master will continously monitor the cluster to verify that all nodes contain a consistant configuration and view of the cluster and will initiate a recovery process when required.

During the recovery phase, the recovery master will automatically rebuild/recovery all clustered TDB database to ensure that the database are consistent. Recovery typically takes between 1 and 3 seconds. During the recovery period the database are frozen, and all database IO operation by CTDB clients are suspend.

To speed up the process of IP takeover and when clients attached to a failed node recovery as fast as possible. CTDB will automatically generate gratutions ARP packets to inform all nodes of the changed MAC address for that IP. CTDB will also send ‘tickle ack’ packets to all attached clients to trigger that clients to immediatlly recognized that the TCP connection needs to be re-established and to shortcut any TCP restransmission time outs that may be active in the clients.

CTDB current Architecture:
1. Multi-process server:
Multiple smbd daemons, one per client
Each daemon attached to a number of database
Database store all shared meta_data
2. Clustering this should be easy:
Why not just use a cluster database
Each smbd talkes to cluster database instead of local database
Obvious solution can be wrong.

CTDB is similarin conceps to Berkeley DB.