This document describes the state of Postgres support in MediaWiki. == Overview == Support for PostgreSQL has been available since version 1.7 of MediaWiki, and is fairly well maintained. The main code is very well integrated, while extensions are very hit and miss. Still, it is probably the most supported database after MySQL. Much of the work in making MediaWiki database-agnostic came about through the work of creating Postgres support. == Required versions == The current minimum version of PostgreSQL for MediaWiki is 8.1. It is expected that this will be raised to 8.3 at some point, as 8.1 and 8.2 are nearing end of life. == Database schema == Postgres has its own schema file at maintenance/postgres/tables.sql. The goal is to keep this file as close as possible to the canonical schema at maintenance/tables.sql, but without copying over all the usage comments. General notes on the conversion: * The use of a true TIMESTAMP rather than the text string that MySQL uses is highly encouraged. There are still places in the code (especially extensions) which make assumptions about the textual nature of timestamp fields, but these can almost always be programmed around. * Although Postgres has a true BOOLEAN type, boolean columns are always mapped to SMALLINT, as the code does not always treat the column as a boolean (which is limited to accepting true, false, 0, 1, t, or f) * The default data type for all VARCHAR, CHAR, and VARBINARY columns should simply be TEXT. The only exception is when VARBINARY is used to store true binary data, such as the math_inputhash column, in which case BYTEA should be used. * All integer variants should generally be mapped to INTEGER. There is small-to-no advantage in using SMALLINT versus INTEGER in Postgres, and the possibility of running out of room outweighs such concerns. The columns that are BIGINT in other schemas should be INTEGER as well, as none of them so far are even remotely likely to reach the 32 billion limit of an INTEGER. * Blobs (blob, tinyblog, mediumblob) should be mapped to TEXT whenever possible, and to BYTEA if they are known to contain binary data. * All length modifiers on data types should be removed. If they are on an INTEGER, it's probably an error, and if on any text-based field, simply using TEXT is preferred. * Sequences should be explicitly named rather than using SERIAL, as the code can depend on having a specific name. * Foreign keys should be used when possible. This makes things both easier and harder in the code, but most of the major problems have now been overcome. Always add an explicit ON DELETE clause, and consider carefully what choice to use (all things considered, prefer CASCADE). * The use of CIDR should be done very carefully, because the code will sometimes want to store things such as an empty string or other non-IP value in the column. When in doubt, use TEXT. * Indexes should be created using the original MySQL tables.sql as a guide, but keeping in mind the ability of Postgres to use partial indexes, functional indexes, and bitmaps. The index names should be logical but are not too important, as they are never referenced directly by the code (unlike sequence names). Most of the indexes in the file as of this writing are there due to production testing of expensive queries on a busy wiki. == Keeping in sync with tables.sql == The script maintenance/postgres/compare_schemas.pl should be periodically run. It will parse both "tables.sql" files and produce any differences found. Such differences should be fixed or exceptions specifically carved out by editing the script itself. This script has also been very useful in finding problems in maintenance/tables.sql itself, as it is very strict in the format it expects things to be in. :) == Getting help == In addition to the normal venues (MediaWiki mailing lists and IRC channels), the #postgresql channel on irc.freenode.net is a friendly and expert resource if you should encounter a problem with your Postgres-enabled MediaWiki.