MediaWiki extension: CirrusSearch

Installation

Get Elasticsearch up and running somewhere. Only Elasticsearch v7.10 is supported. A compatability layer for writing to Elasticsearch v6.8 is provided to support zero-downtime migrations when multiple clusters are available. If you will be running Elasticsearch on a host separate from the mediawiki installation see https://www.elastic.co/guide/en/elasticsearch/reference/current/modules-network.html. Be careful with the network configuration, never expose an unprotected node to the internet.

Place the CirrusSearch extension in your extensions directory. You also need to install the Elastica MediaWiki extension. Add this to LocalSettings.php: wfLoadExtension( 'Elastica' ); wfLoadExtension( 'CirrusSearch' ); $wgDisableSearchUpdate = true;

Configure your search servers in LocalSettings.php if you aren't running Elasticsearch on localhost: $wgCirrusSearchServers = [ 'elasticsearch0', 'elasticsearch1', 'elasticsearch2', 'elasticsearch3' ]; There are other $wgCirrusSearch variables that you might want to change from their defaults.

Now run this script to generate your elasticsearch index: php $MW_INSTALL_PATH/extensions/CirrusSearch/maintenance/UpdateSearchIndexConfig.php

Now remove $wgDisableSearchUpdate = true from LocalSettings.php. Updates should start heading to Elasticsearch.

Next bootstrap the search index by running: php $MW_INSTALL_PATH/extensions/CirrusSearch/maintenance/ForceSearchIndex.php --skipLinks --indexOnSkip php $MW_INSTALL_PATH/extensions/CirrusSearch/maintenance/ForceSearchIndex.php --skipParse Note that this can take some time. For large wikis read "Bootstrapping large wikis" below.

Once that is complete add this to LocalSettings.php to funnel queries to ElasticSearch: $wgSearchType = 'CirrusSearch';

Bootstrapping large wikis

Since most of the load involved in indexing is parsing the pages in php we provide a few options to split the process into multiple processes. Don't worry too much about the database during this process. It can generally handle more indexing processes then you are likely to be able to spawn.

General strategy: 0. Make sure you have a good job queue setup. It'll be doing most of the work. In fact, Cirrus won't work well on large wikis without it.

1. Generate scripts to add all the pages without link counts to the index.
2. Execute them any way you like.
3. Generate scripts to count all the links.
4. Execute them any way you like.

Step 1: In bash I do this: export PROCS=5 #or whatever number you want rm -rf cirrus_scripts mkdir cirrus_scripts mkdir cirrus_log pushd cirrus_scripts php extensions/CirrusSearch/maintenance/ForceSearchIndex.php --queue --maxJobs 10000 --pauseForJobs 1000 \

--skipLinks --indexOnSkip --buildChunks 250000 |
sed -e 's/$/ | tee -a cirrus_log\/'$wiki'.parse.log/' |
split -n r/$PROCS

for script in x*; do sort -R $script > $script.sh && rm $script; done popd

Step 2: Just run all the scripts that step 1 made. Best to run them in screen or something and in the directory above cirrus_scripts. So like this: bash cirrus_scripts/xaa.sh

Step 3: In bash I do this: pushd cirrus_scripts rm *.sh php extensions/CirrusSearch/maintenance/ForceSearchIndex.php --queue --maxJobs 10000 --pauseForJobs 1000 \

--skipParse --buildChunks 250000 |
sed -e 's/$/ | tee -a cirrus_log\/'$wiki'.parse.log/' |
split -n r/$PROCS

for script in x*; do sort -R $script > $script.sh && rm $script; done popd

Step 4: Same as step 2 but for the new scripts. These scripts put more load on Elasticsearch so you might want to run them just one at a time if you don't have a huge Elasticsearch cluster or you want to make sure not to cause load spikes.

If you don't have a good job queue you can try the above but lower the buildChunks parameter significantly and remove the --queue parameter.

Handling elasticsearch outages

If for some reason in process updates to elasticsearch begin failing you can immediately set "$wgDisableSearchUpdate = true;" in your LocalSettings.php file to stop trying to update elasticsearch. Once you figure out what is wrong with elasticsearch you should turn those updates back on and then run the following: php ./maintenance/ForceSearchIndex.php --from <whenever the outage started in ISO 8601 format> --deletes php ./maintenance/ForceSearchIndex.php --from <whenever the outage started in ISO 8601 format>

The first command picks up all the deletes that occurred during the outage and should complete quite quickly. The second command picks up all the updates that occurred during the outage and might take significantly longer.

Changing $wgNamespacesToBeSearchedDefault

When changing wgNamespacesToBeSearchedDefault you might need to reindex some pages from the source documents. For achieving this you have two options:

• Blow away the search index and rebuild it from scratch. (see the Upgrading section, option A)
• Use the "Saneitizer": php Saneitize.php
  
  Both options have drawbacks, the first one might incur a downtime making some pages unfindable while they are indexed, the second option might take time if the wiki is large.

PoolCounter

CirrusSearch can leverage the PoolCounter extension to limit the number of concurrent searches to elasticsearch. You can do this by installing the PoolCounter extension and then configuring it in LocalSettings.php like so: wfLoadExtension( 'PoolCounter'); Configuration for standard searches. $wgPoolCounterConf[ 'CirrusSearch-Search' ] = [ 'class' => 'MediaWiki\PoolCounter\PoolCounterClient', 'timeout' => 30, 'workers' => 25, 'maxqueue' => 50, ]; Configuration for prefix searches. These are usually quite quick and plentiful. $wgPoolCounterConf[ 'CirrusSearch-Prefix' ] = [ 'class' => 'MediaWiki\PoolCounter\PoolCounterClient', 'timeout' => 10, 'workers' => 50, 'maxqueue' => 100, ]; Configuration for regex searches. These are slow and use lots of resources so we only allow a few at a time. $wgPoolCounterConf[ 'CirrusSearch-Regex' ] = [ 'class' => 'MediaWiki\PoolCounter\PoolCounterClient', 'timeout' => 30, 'workers' => 10, 'maxqueue' => 10, ]; Configuration for funky namespace lookups. These should be reasonably fast // and reasonably rare. $wgPoolCounterConf[ 'CirrusSearch-NamespaceLookup' ] = [

		'class' => 'MediaWiki\PoolCounter\PoolCounterClient',
		'timeout' => 10,
		'workers' => 20,
		'maxqueue' => 20,

), ];

Upgrading

When you upgrade there four possible cases for maintaining the index:

1. You must update the index configuration and reindex from source documents.
2. You must update the index configuration and reindex from already indexed documents.
3. You must update the index configuration but no reindex is required.
4. No changes are required.

If you must do (1) you have two options: A. Blow away the search index and rebuild it from scratch. Marginally faster and uses less disk space on in elasticsearch but empties the index entirely and rebuilds it so search will be down for a while: php updateSearchIndexConfig.php --startOver php forceSearchIndex.php

B. Build a copy of the index, reindex to it, and then force a full reindex from source documents. Uses more disk space but search should be up the entire time: php updateSearchIndexConfig.php --reindexAndRemoveOk --indexIdentifier now php forceSearchIndex.php

If you must do (2) really have only one option: A. Build of a copy of the index and reindex to it: php updateSearchIndexConfig.php --reindexAndRemoveOk --indexIdentifier now php forceSearchIndex.php --from <time when you started updateSearchIndexConfig.php in YYYY-mm-ddTHH:mm:ssZ> --deletes php forceSearchIndex.php --from <time when you started updateSearchIndexConfig.php in YYYY-mm-ddTHH:mm:ssZ> or for the Bash inclined: TZ=UTC export REINDEX_START=$(date +%Y-%m-%dT%H:%m:%SZ) php updateSearchIndexConfig.php --reindexAndRemoveOk --indexIdentifier now php forceSearchIndex.php --from $REINDEX_START --deletes php forceSearchIndex.php --from $REINDEX_START

If you must do (3) you again only have one option: A. Same as (2.A)

4 is easy!

The safest thing if you don't know what is required for your update is to execute (1.B).

Production suggestions

Elasticsearch

All the general rules for making Elasticsearch production ready apply here. So you don't have to go round them up below is a list. Some of these steps are obvious, others will take some research.

• NOTE: this list was written for 0.90 so it may not work well for 1.0. It'll be revised when I have

more experience with 1.0. --Nik

1. Have >= 3 nodes.
2. Configure Elasticsearch for memlock.
3. Change each node's elasticsearch.yml file in a few ways.

3a. Change node name to the real host name. 3b. Turn off auto creation and some other scary stuff by adding this (tested with 0.90.4):

1. 1. 1. Actions #########
   2. Modulo some small changes to comments this section comes directly from the
   3. wonderful Elasticsearch mailing list, specifically Dan Everton. ##
2. Require explicit index creation. ES never auto creates the indexes the way we
3. like them. ## action.auto_create_index: false
   
   ##
4. Protect against accidental close/delete operations on all indices. You can
5. still close/delete individual indices. ## action.disable_close_all_indices: true action.disable_delete_all_indices: true
   
   ##
6. Disable ability to shutdown nodes via REST API. ## action.disable_shutdown: true

Testing

See tests

Job Queue

Cirrus makes heavy use of the job queue. You can run it without any job queue customization but if you switch the job queue to Redis with checkDelay enabled then Cirrus's results will be more correct. The reason for this is that this configuration allows Cirrus to delay link counts until Elasticsearch has appropriately refreshed. This is an example of configuring it: $redisPassword = '<password goes here>'; $wgJobTypeConf['default'] = [ 'class' => 'JobQueueRedis', 'order' => 'fifo', 'redisServer' => 'localhost', 'checkDelay' => true, 'redisConfig' => [

		'password' => $redisPassword,

], ];

Note: some MediaWiki setups have trouble running the job queue. It can be finicky. The most sure fire way to get it to work is also the slowest. Add this to your LocalSettings.php: $wgRunJobsAsync = false;

Development

The fastest way to get started with CirrusSearch development is to use MediaWiki-Vagrant.

1. Follow steps here: https://www.mediawiki.org/wiki/MediaWiki-Vagrant#Quick_start
2. Now execute the following:

vagrant enable-role cirrussearch vagrant provision

This can take some time but it produces a clean development environment in a virtual machine that has everything required to run Cirrus.

Hooks

See docs/hooks.txt.

Licensing information

CirrusSearch makes use of the Elastica extension containing the Elastica library to connect to Elasticsearch http://elastica.io/. It is Apache licensed and you can read the license Elastica/LICENSE.txt.