LoadMonitor.php

Go to the documentation of this file.

<?php
namespace Wikimedia\Rdbms;
 
use BagOStuff;
use Liuggio\StatsdClient\Factory\StatsdDataFactoryInterface;
use NullStatsdDataFactory;
use Psr\Log\LoggerInterface;
use Psr\Log\NullLogger;
use RuntimeException;
use WANObjectCache;
use Wikimedia\ScopedCallback;
 
class LoadMonitor implements ILoadMonitor {
    protected $lb;
    protected $srvCache;
    protected $wanCache;
    protected $replLogger;
    protected $statsd;
 
    private $movingAveRatio;
    private $lagWarnThreshold;
 
    private $wallClockOverride;
 
    private $serverStatesKeyLocked = false;
 
    private const VERSION = 1;
    private const POLL_PERIOD_MS = 500;
    private const STATE_PRESERVE_TTL = 60;
    private const TIME_TILL_REFRESH = 1;
 
    public function __construct(
        ILoadBalancer $lb, BagOStuff $srvCache, WANObjectCache $wCache, array $options = []
    ) {
        $this->lb = $lb;
        $this->srvCache = $srvCache;
        $this->wanCache = $wCache;
        $this->replLogger = new NullLogger();
        $this->statsd = new NullStatsdDataFactory();
 
        $this->movingAveRatio = $options['movingAveRatio'] ?? 0.1;
        $this->lagWarnThreshold = $options['lagWarnThreshold'] ?? LoadBalancer::MAX_LAG_DEFAULT;
    }
 
    public function setLogger( LoggerInterface $logger ) {
        $this->replLogger = $logger;
    }
 
    public function setStatsdDataFactory( StatsdDataFactoryInterface $statsFactory ) {
        $this->statsd = $statsFactory;
    }
 
    final public function scaleLoads( array &$weightByServer, $domain ) {
        $serverIndexes = array_keys( $weightByServer );
        $states = $this->getServerStates( $serverIndexes, $domain );
        $newScalesByServer = $states['weightScales'];
        foreach ( $weightByServer as $i => $weight ) {
            if ( isset( $newScalesByServer[$i] ) ) {
                $weightByServer[$i] = (int)ceil( $weight * $newScalesByServer[$i] );
            } else { // server recently added to config?
                $host = $this->lb->getServerName( $i );
                $this->replLogger->error( __METHOD__ . ": host $host not in cache" );
            }
        }
    }
 
    final public function getLagTimes( array $serverIndexes, $domain ) {
        return $this->getServerStates( $serverIndexes, $domain )['lagTimes'];
    }
 
    protected function getServerStates( array $serverIndexes, $domain ) {
        // Represent the cluster by the name of the primary DB
        $cluster = $this->lb->getServerName( $this->lb->getWriterIndex() );
 
        // Randomize logical TTLs to reduce stampedes
        $ageStaleSec = mt_rand( 1, self::POLL_PERIOD_MS ) / 1e3;
        $minAsOfTime = $this->getCurrentTime() - $ageStaleSec;
 
        // (a) Check the local server cache
        $srvCacheKey = $this->getStatesCacheKey( $this->srvCache, $serverIndexes );
        $value = $this->srvCache->get( $srvCacheKey );
        if ( $value && $value['timestamp'] > $minAsOfTime ) {
            $this->replLogger->debug( __METHOD__ . ": used fresh '$cluster' cluster status" );
 
            return $value; // cache hit
        }
 
        // (b) Value is stale/missing; try to use/refresh the shared cache
        $scopedLock = $this->srvCache->getScopedLock( $srvCacheKey, 0, 10 );
        if ( !$scopedLock && $value ) {
            $this->replLogger->debug( __METHOD__ . ": used stale '$cluster' cluster status" );
            // (b1) Another thread on this server is already checking the shared cache
            return $value;
        }
 
        // (b2) This thread gets to check the shared cache or (b3) value is missing
        $staleValue = $value;
        $updated = false; // whether the regeneration callback ran
        $value = $this->wanCache->getWithSetCallback(
            $this->getStatesCacheKey( $this->wanCache, $serverIndexes ),
            self::TIME_TILL_REFRESH, // 1 second logical expiry
            function ( $oldValue, &$ttl ) use ( $serverIndexes, $domain, $staleValue, &$updated ) {
                // Double check for circular recursion in computeServerStates()/getWeightScale().
                // Mainly, connection attempts should use LoadBalancer::getServerConnection()
                // rather than something that will pick a server based on the server states.
                $scopedLock = $this->acquireServerStatesLoopGuard();
                if ( !$scopedLock ) {
                    throw new RuntimeException(
                        "Circular recursion detected while regenerating server states cache. " .
                        "This may indicate improper connection handling in " . get_class( $this )
                    );
                }
 
                $updated = true;
 
                return $this->computeServerStates(
                    $serverIndexes,
                    $domain,
                    $oldValue ?: $staleValue // fallback to local cache stale value
                );
            },
            [
                // One thread can update at a time; others use the old value
                'lockTSE' => self::STATE_PRESERVE_TTL,
                'staleTTL' => self::STATE_PRESERVE_TTL,
                // If there is no shared stale value then use the local cache stale value;
                // When even that is not possible, then use the trivial value below.
                'busyValue' => $staleValue ?: $this->getPlaceholderServerStates( $serverIndexes )
            ]
        );
 
        if ( $updated ) {
            $this->replLogger->info( __METHOD__ . ": regenerated '$cluster' cluster status" );
        } else {
            $this->replLogger->debug( __METHOD__ . ": used cached '$cluster' cluster status" );
        }
 
        // Backfill the local server cache
        if ( $scopedLock ) {
            $this->srvCache->set( $srvCacheKey, $value, self::STATE_PRESERVE_TTL );
        }
 
        return $value;
    }
 
    protected function computeServerStates( array $serverIndexes, $domain, $priorStates ) {
        // Check if there is just a primary DB (no replication involved)
        if ( $this->lb->getServerCount() <= 1 ) {
            return $this->getPlaceholderServerStates( $serverIndexes );
        }
 
        $priorScales = $priorStates ? $priorStates['weightScales'] : [];
        $cluster = $this->lb->getClusterName();
 
        $lagTimes = [];
        $weightScales = [];
        foreach ( $serverIndexes as $i ) {
            $isPrimary = ( $i == $this->lb->getWriterIndex() );
            // If the primary DB has zero load, then typical read queries do not use it.
            // In that case, avoid connecting to it since this method might run in any
            // datacenter, and the primary DB might be geographically remote.
            if ( $isPrimary && $this->lb->getServerInfo( $i )['load'] <= 0 ) {
                $lagTimes[$i] = 0;
                // Callers only use this DB if they have *no choice* anyway (e.g. writes)
                $weightScales[$i] = 1.0;
                continue;
            }
 
            $host = $this->lb->getServerName( $i );
            # Handles with open transactions are avoided since they might be subject
            # to REPEATABLE-READ snapshots, which could affect the lag estimate query.
            $flags = ILoadBalancer::CONN_TRX_AUTOCOMMIT | ILoadBalancer::CONN_SILENCE_ERRORS;
            $conn = $this->lb->getAnyOpenConnection( $i, $flags );
            if ( $conn ) {
                $close = false; // already open
            } else {
                // Get a connection to this server without triggering other server connections
                $conn = $this->lb->getServerConnection( $i, ILoadBalancer::DOMAIN_ANY, $flags );
                $close = true; // new connection
            }
 
            // Get new weight scale using a moving average of the naïve and prior values
            $lastScale = $priorScales[$i] ?? 1.0;
            $naiveScale = $this->getWeightScale( $i, $conn ?: null );
            $newScale = $this->getNewScaleViaMovingAve(
                $lastScale,
                $naiveScale,
                $this->movingAveRatio
            );
            // Scale from 0% to 100% of nominal weight
            $newScale = max( $newScale, 0.0 );
 
            $weightScales[$i] = $newScale;
            $statHost = str_replace( '.', '_', $host );
            $this->statsd->gauge( "loadbalancer.weight.$cluster.$statHost", $newScale );
 
            // Mark replication lag on this server as "false" if it is unreachable
            if ( !$conn ) {
                $lagTimes[$i] = $isPrimary ? 0 : false;
                $this->replLogger->error(
                    __METHOD__ . ": host {db_server} is unreachable",
                    [ 'db_server' => $host ]
                );
                continue;
            }
 
            // Determine the amount of replication lag on this server
            try {
                $lag = $conn->getLag();
            } catch ( DBError $e ) {
                // Mark the lag time as "false" if it cannot be queried
                $lag = false;
            }
            $lagTimes[$i] = $lag;
 
            if ( $lag === false ) {
                $this->replLogger->error(
                    __METHOD__ . ": host {db_server} is not replicating?",
                    [ 'db_server' => $host ]
                );
            } else {
                $this->statsd->timing( "loadbalancer.lag.$cluster.$statHost", $lag * 1000 );
                if ( $lag > $this->lagWarnThreshold ) {
                    $this->replLogger->warning(
                        "Server {db_server} has {lag} seconds of lag (>= {maxlag})",
                        [
                            'db_server' => $host,
                            'lag' => $lag,
                            'maxlag' => $this->lagWarnThreshold
                        ]
                    );
                }
            }
 
            if ( $close ) {
                # Close the connection to avoid sleeper connections piling up.
                # Note that the caller will pick one of these DBs and reconnect,
                # which is slightly inefficient, but this only matters for the lag
                # time cache miss cache, which is far less common that cache hits.
                $this->lb->closeConnection( $conn );
            }
        }
 
        return [
            'lagTimes' => $lagTimes,
            'weightScales' => $weightScales,
            'timestamp' => $this->getCurrentTime()
        ];
    }
 
    private function getPlaceholderServerStates( array $serverIndexes ) {
        return [
            'lagTimes' => array_fill_keys( $serverIndexes, 0 ),
            'weightScales' => array_fill_keys( $serverIndexes, 1.0 ),
            'timestamp' => $this->getCurrentTime()
        ];
    }
 
    protected function getWeightScale( $index, IDatabase $conn = null ) {
        return $conn ? 1.0 : 0.0;
    }
 
    protected function getNewScaleViaMovingAve( $lastScale, $naiveScale, $movAveRatio ) {
        return $movAveRatio * $naiveScale + ( 1 - $movAveRatio ) * $lastScale;
    }
 
    private function getStatesCacheKey( $cache, array $serverIndexes ) {
        sort( $serverIndexes );
        // Lag is per-server, not per-DB, so key on the primary DB name
        return $cache->makeGlobalKey(
            'rdbms-server-states',
            self::VERSION,
            $this->lb->getServerName( $this->lb->getWriterIndex() ),
            implode( '-', $serverIndexes )
        );
    }
 
    private function acquireServerStatesLoopGuard() {
        if ( $this->serverStatesKeyLocked ) {
            return null; // locked
        }
 
        $this->serverStatesKeyLocked = true; // lock
 
        return new ScopedCallback( function () {
            $this->serverStatesKeyLocked = false; // unlock
        } );
    }
 
    protected function getCurrentTime() {
        return $this->wallClockOverride ?: microtime( true );
    }
 
    public function setMockTime( &$time ) {
        $this->wallClockOverride =& $time;
    }
}