Elvis is a Postgresql replication project with several years of intense production usage in several different contexts. It is uses triggers to put changes into a change table and daemons to apply those changes to slave databases. Transactional consistency is maintained at a granularity you control.

It consists of a c plugin, some pgsql functions, and a java daemon. The c plugin knows how to take changes and add them to the change table, and how to apply a change to a slave table.

We liked the flexible per table trigger nature of slony, but didn't like the message complexity of Slony which is O((Nodes * Changes)^2) since all Slony cluster members are required to know the status of every other member.

Replication is asynchronous but respects transactions; slaves recieve changes in batches that correspond to groups of one or more transactions from the master. The granularity of the groups of transaction is determined by a snapshot table (more below).

The replication topology for a given table should be a tree with changes arising only at the root (more below).

To make a table the master you apply the master trigger. To make it a slave you apply the slave trigger. If a table has both triggers it is a relay table.

You can flexibly mix and match which tables are masters and slaves on which machines.

The c function encodes changes into before and after strings and puts them into the change table along with the current transaction id.

The daemon creates periodic transactions and notes their transaction ids into a snapshot table.

The tree topology is stored in a special schema, a copy of which must be available to daemons when they startup.

One daemon per master or relay node will have a thread for each slave; each slave selects its changes from the database separately so they won't hold each other up. Changes are cached within the daemon to reduce load on the master for the situation of many slaves.

Slave threads update slavesnapshotstatus on the slave database to mark their progress.

see javasrc/com/netblue/bruce/SlaveRunner.java, starting from doWork().

see schema/replication-helpers.sql and schema/replication-ddl.sql

• It does not handle transaction id rollover, requiring complex manual intervention when your databse XID, an unsigned 32bit integer, cycles across 2^32 and goes back to zero.

• slavesnapshotstatus requires frequent vacuuming. See the postgresql vacuum command documentation.

• There is no documentation or support.

• It is possible to change the replication topology on the fly but it is tricky and requires a bit of planning.

• The triggers enforce a requirement for a primary key; this protects against unintential UPDATE and DELETE but there are some valid INSERT only scenarios that would work if this wasn't enforced.

• PUBLIC can insert changes into the transactionlog and snapshots table; you could manually narrow this to all users who need to be able to make changes. This drawback is due to the fact that all changes are comingled in one table; that design decision is driven by concerns over the complexity of alternatives. For instance one possible approach to this is to maintain a separate permissions table and use a trigger to validate that at insert time, but note this check would be applied for each row changed.

• When a change is applied to a row the before images must match or an error is thrown and replication to the node stops. There should be more flexible ways to deal with issues that arise from rare instances of corruption without magnifying the impact. More flexible conflict resolution policies would also enable multi master.

• The tree topology should be specified using yaml or a custom config file format, or a simply as command line options to the individual daemons; this would remove the single point of failure of the schema node.

Elvis saw several years of constant high volume production use with Postgresql 8.0, 8.1, and 8.3. There is a an issue on 8.2 that prevents it being used for production. It hasn't been tested on 9.x or developed or maintained for about two years.

Elvis was sponsored by Connexus, which was purchased by Epic Media in 2010.

Elvis is based on Slony. See slon2_20070108.txt for original design notes by Bob Klahn.

Elvis used to be named Bruce after the monty python skit in which everyone is named Bruce because its easier that way. It was renamed to Elvis because "Elvis has left the building".

The first developer was Bob Klahn who wrote the first version of the c plugin. Lance Ball wrote early versions of the java daemon. Matt Grosso provided early kibitzing. Further development needed to make it stable in production was done mostly by Matt Grosso, with important contributions from Kalpesh Patel and Abhay Saswade.