What is Certification based Replication?
Certification based replication uses group communication and transaction ordering techniques to achieve synchronous replication. Transactions execute optimistically in a single node (or replica) and, at commit time, run a coordinated certification process to enforce global consistency. Global coordination is achieved with the help of a broadcast service, that establishes a global total order among concurrent transactions.
Pre-requisites for certification based replication:
- database is transactional (i.e. it can rollback uncommitted changes)
- each replication event changes the database atomically
- replicated events are globally ordered (i.e. applied on all instances in the same order)
The main idea is that a transaction is executed conventionally until the commit point, under the assumption that there will be no conflict. This is called optimistic execution. When the client issues a COMMIT command (but before the actual commit has happened), all changes made to the database by the transaction and the primary keys of changed rows are collected into a writeset. This writeset is then replicated to the rest of the nodes. After that, the writeset undergoes a deterministic certification test (using the collected primary keys) on each node (including the writeset originator node) which determines if the writeset can be applied or not.
If the certification test fails, the writeset is dropped and the original transaction is rolled back. If the test succeeds, the transaction is committed and the writeset is applied on the rest of the nodes.
The certification test implemented in Galera depends on the global ordering of transactions. Each transaction is assigned a global ordinal sequence number during replication. Thus, when a transaction reaches the commit point, it is known what was the sequence number of the last transaction it did not conflict with. The interval between those two numbers is an uncertainty land: transactions in this interval have not seen the effects of each other. Therefore, all transactions in this interval are checked for primary key conflicts with the transaction in question. The certification test fails if a conflict is detected.
Since the procedure is deterministic and all replicas receive transactions in the same order, all nodes reach the same decision about the outcome of the transaction. The node that started the transaction can then notify the client application if the transaction has been committed or not.
Certification based replication (or more precisely, certification-based conflict resolution) is based on academic research, in particular on Fernando Pedone’s Ph.D. thesis (http://www.inf.usi.ch/faculty/pedone/Paper/199x/These-2090-Pedone.pdf).
Please sign in to leave a comment.
Comments
0 comments