Current File : /www/varak.net/wiki.varak.net/extensions/CirrusSearch/includes/Elastica/PooledHttp.php
<?php
namespace CirrusSearch\Elastica;
use Elastica\Transport\Http;
use MediaWiki\Logger\LoggerFactory;
use MediaWiki\MediaWikiServices;
use RequestContext;
/**
* Implements cross-request connection pooling via hhvm's built in
* curl_init_pooled. To utilize this transport $wgCirrusSearchClusters should
* be configured as follows:
*
* $wgCirrusSearchClusters = array(
* 'default' => array(
* 'transport' => 'CirrusSearch\Elastica\PooledHttp',
* 'port' => 12345,
* 'host' => 'my.host.name',
* 'config' => array( 'pool' => 'cirrus' ),
* )
* );
*
* Additionally hhvm needs the following configuration in its php.ini:
*
* curl.namedPools = cirrus
*
* Other optional pool parameters (also in php.ini) and their defaults:
*
* curl.namedPools.cirrus.size = 5
* curl.namedPools.cirrus.reuseLimit = 100
* curl.namedPools.cirrus.connGetTimeout = 5000
*
* For connection pooling to work optimally you will want to configure a pool
* for each host you connect to. This means using a different pool for each
* cluster, and using some sort of load balancer that allows to connect to
* your entire cluster using a single ip or domain name.
*/
class PooledHttp extends Http {
/**
* @param bool $persistent
* @return resource Curl handle
*/
protected function _getConnection( $persistent = true ) {
$conn = $this->getConnection();
if ( !$persistent ) {
return parent::_getConnection( $persistent );
}
$ch = null;
if ( !$conn->hasConfig( 'pool' ) ) {
LoggerFactory::getInstance( 'CirrusSearch' )->warning(
"Elastica connection pool cannot init: missing pool name in connection config"
);
} elseif ( !function_exists( 'curl_init_pooled' ) ) {
LoggerFactory::getInstance( 'CirrusSearch' )->warning(
"Elastica connection pool cannot init: missing curl_init_pooled method. Are you using hhvm >= 3.9.0?"
);
} else {
$pool = $conn->getConfig( 'pool' );
// Note that if the connection pool is full this will block
// up to curl.namedPools.$pool.connGetTimeout ms, defaulting
// to 5000. If the timeout is reached hhvm will raise a fatal
// error.
$start = microtime( true );
$ch = curl_init_pooled( $pool );
$this->reportTiming( microtime( true ) - $start );
if ( $ch === null ) {
LoggerFactory::getInstance( 'CirrusSearch' )->warning(
"Elastic connection pool cannot init: Unknown pool {pool}. Did you configure curl.namedPools?",
[ 'pool' => $pool ]
);
}
}
if ( $ch === null ) {
return parent::_getConnection( $persistent );
}
return $ch;
}
/**
* @param float $tookS The number of seconds taken to get the pooled handle
*/
protected function reportTiming( $tookS ) {
// Should this in some way collect stats about requests that take longer than normal,
// or just allow standard p95 or p99 to collect them? With thousands of requests per
// second i wonder if a couple servers having issues with their pools will by lost in
// the noise. Another option could be recording per-server stats and including gethostname()
// in the key?
MediaWikiServices::getInstance()->getStatsdDataFactory()->timing(
'CirrusSearch.connectionPool.initMs',
intval( 1000 * $tookS )
);
}
}
Mini Shell