Some information about database access in MediaWiki. By Tim Starling, January 2006. ------------------------------------------------------------------------ Database layout ------------------------------------------------------------------------ For information about the MediaWiki database layout, such as a description of the tables and their contents, please see: http://www.mediawiki.org/wiki/Manual:Database_layout http://svn.wikimedia.org/viewvc/mediawiki/trunk/phase3/maintenance/tables.sql?view=markup ------------------------------------------------------------------------ API ------------------------------------------------------------------------ To make a read query, something like this usually suffices: $dbr = wfGetDB( DB_SLAVE ); $res = $dbr->select( /* ...see docs... */ ); while ( $row = $dbr->fetchObject( $res ) ) { ... } $dbr->freeResult( $res ); Note the assignment operator in the while condition. For a write query, use something like: $dbw = wfGetDB( DB_MASTER ); $dbw->insert( /* ...see docs... */ ); We use the convention $dbr for read and $dbw for write to help you keep track of whether the database object is a slave (read-only) or a master (read/write). If you write to a slave, the world will explode. Or to be precise, a subsequent write query which succeeded on the master may fail when replicated to the slave due to a unique key collision. Replication on the slave will stop and it may take hours to repair the database and get it back online. Setting read_only in my.cnf on the slave will avoid this scenario, but given the dire consequences, we prefer to have as many checks as possible. We provide a query() function for raw SQL, but the wrapper functions like select() and insert() are usually more convenient. They take care of things like table prefixes and escaping for you. If you really need to make your own SQL, please read the documentation for tableName() and addQuotes(). You will need both of them. ------------------------------------------------------------------------ Basic query optimisation ------------------------------------------------------------------------ MediaWiki developers who need to write DB queries should have some understanding of databases and the performance issues associated with them. Patches containing unacceptably slow features will not be accepted. Unindexed queries are generally not welcome in MediaWiki, except in special pages derived from QueryPage. It's a common pitfall for new developers to submit code containing SQL queries which examine huge numbers of rows. Remember that COUNT(*) is O(N), counting rows in a table is like counting beans in a bucket. ------------------------------------------------------------------------ Replication ------------------------------------------------------------------------ The largest installation of MediaWiki, Wikimedia, uses a large set of slave MySQL servers replicating writes made to a master MySQL server. It is important to understand the issues associated with this setup if you want to write code destined for Wikipedia. It's often the case that the best algorithm to use for a given task depends on whether or not replication is in use. Due to our unabashed Wikipedia-centrism, we often just use the replication-friendly version, but if you like, you can use $wgLoadBalancer->getServerCount() > 1 to check to see if replication is in use. === Lag === Lag primarily occurs when large write queries are sent to the master. Writes on the master are executed in parallel, but they are executed in serial when they are replicated to the slaves. The master writes the query to the binlog when the transaction is committed. The slaves poll the binlog and start executing the query as soon as it appears. They can service reads while they are performing a write query, but will not read anything more from the binlog and thus will perform no more writes. This means that if the write query runs for a long time, the slaves will lag behind the master for the time it takes for the write query to complete. Lag can be exacerbated by high read load. MediaWiki's load balancer will stop sending reads to a slave when it is lagged by more than 30 seconds. If the load ratios are set incorrectly, or if there is too much load generally, this may lead to a slave permanently hovering around 30 seconds lag. If all slaves are lagged by more than 30 seconds, MediaWiki will stop writing to the database. All edits and other write operations will be refused, with an error returned to the user. This gives the slaves a chance to catch up. Before we had this mechanism, the slaves would regularly lag by several minutes, making review of recent edits difficult. In addition to this, MediaWiki attempts to ensure that the user sees events occurring on the wiki in chronological order. A few seconds of lag can be tolerated, as long as the user sees a consistent picture from subsequent requests. This is done by saving the master binlog position in the session, and then at the start of each request, waiting for the slave to catch up to that position before doing any reads from it. If this wait times out, reads are allowed anyway, but the request is considered to be in "lagged slave mode". Lagged slave mode can be checked by calling $wgLoadBalancer->getLaggedSlaveMode(). The only practical consequence at present is a warning displayed in the page footer. === Lag avoidance === To avoid excessive lag, queries which write large numbers of rows should be split up, generally to write one row at a time. Multi-row INSERT ... SELECT queries are the worst offenders should be avoided altogether. Instead do the select first and then the insert. === Working with lag === Despite our best efforts, it's not practical to guarantee a low-lag environment. Lag will usually be less than one second, but may occasionally be up to 30 seconds. For scalability, it's very important to keep load on the master low, so simply sending all your queries to the master is not the answer. So when you have a genuine need for up-to-date data, the following approach is advised: 1) Do a quick query to the master for a sequence number or timestamp 2) Run the full query on the slave and check if it matches the data you got from the master 3) If it doesn't, run the full query on the master To avoid swamping the master every time the slaves lag, use of this approach should be kept to a minimum. In most cases you should just read from the slave and let the user deal with the delay. ------------------------------------------------------------------------ Lock contention ------------------------------------------------------------------------ Due to the high write rate on Wikipedia (and some other wikis), MediaWiki developers need to be very careful to structure their writes to avoid long-lasting locks. By default, MediaWiki opens a transaction at the first query, and commits it before the output is sent. Locks will be held from the time when the query is done until the commit. So you can reduce lock time by doing as much processing as possible before you do your write queries. Update operations which do not require database access can be delayed until after the commit by adding an object to $wgPostCommitUpdateList. Often this approach is not good enough, and it becomes necessary to enclose small groups of queries in their own transaction. Use the following syntax: $dbw = wfGetDB( DB_MASTER ); $dbw->immediateBegin(); /* Do queries */ $dbw->immediateCommit(); There are functions called begin() and commit() but they don't do what you would expect. Don't use them. Use of locking reads (e.g. the FOR UPDATE clause) is not advised. They are poorly implemented in InnoDB and will cause regular deadlock errors. It's also surprisingly easy to cripple the wiki with lock contention. If you must use them, define a new flag for $wgAntiLockFlags which allows them to be turned off, because we'll almost certainly need to do so on the Wikimedia cluster. Instead of locking reads, combine your existence checks into your write queries, by using an appropriate condition in the WHERE clause of an UPDATE, or by using unique indexes in combination with INSERT IGNORE. Then use the affected row count to see if the query succeeded.