JobRunner.php

Go to the documentation of this file.

<?php
use Liuggio\StatsdClient\Factory\StatsdDataFactoryInterface;
use MediaWiki\Config\ServiceOptions;
use MediaWiki\Logger\LoggerFactory;
use MediaWiki\MediaWikiServices;
use Psr\Log\LoggerAwareInterface;
use Psr\Log\LoggerInterface;
use Wikimedia\Rdbms\DBError;
use Wikimedia\Rdbms\DBReadOnlyError;
use Wikimedia\Rdbms\ILBFactory;
use Wikimedia\ScopedCallback;
 
class JobRunner implements LoggerAwareInterface {
 
    public const CONSTRUCTOR_OPTIONS = [
        'JobBackoffThrottling',
        'JobClasses',
        'JobSerialCommitThreshold',
        'MaxJobDBWriteDuration',
        'TrxProfilerLimits'
    ];
 
    private $options;
 
    private $lbFactory;
 
    private $jobQueueGroup;
 
    private $readOnlyMode;
 
    private $linkCache;
 
    private $stats;
 
    private $debug;
 
    private $logger;
 
    private const MAX_ALLOWED_LAG = 3;
    private const SYNC_TIMEOUT = self::MAX_ALLOWED_LAG;
    private const LAG_CHECK_PERIOD = 1.0;
    private const ERROR_BACKOFF_TTL = 1;
    private const READONLY_BACKOFF_TTL = 30;
 
    public function setDebugHandler( $debug ) {
        $this->debug = $debug;
    }
 
    public function setLogger( LoggerInterface $logger ) {
        wfDeprecated( __METHOD__, '1.35' );
        $this->logger = $logger;
    }
 
    public function __construct(
        $serviceOptions = null,
        ILBFactory $lbFactory = null,
        JobQueueGroup $jobQueueGroup = null,
        ReadOnlyMode $readOnlyMode = null,
        LinkCache $linkCache = null,
        StatsdDataFactoryInterface $statsdDataFactory = null,
        LoggerInterface $logger = null
    ) {
        if ( !$serviceOptions || $serviceOptions instanceof LoggerInterface ) {
            // TODO: wfDeprecated( __METHOD__ . ' called directly. Use MediaWikiServices instead', '1.35' );
            $logger = $serviceOptions;
            $serviceOptions = new ServiceOptions(
                static::CONSTRUCTOR_OPTIONS,
                MediaWikiServices::getInstance()->getMainConfig()
            );
        }
 
        $this->options = $serviceOptions;
        $this->lbFactory = $lbFactory ?? MediaWikiServices::getInstance()->getDBLoadBalancerFactory();
        $this->jobQueueGroup = $jobQueueGroup ?? JobQueueGroup::singleton();
        $this->readOnlyMode = $readOnlyMode ?: MediaWikiServices::getInstance()->getReadOnlyMode();
        $this->linkCache = $linkCache ?? MediaWikiServices::getInstance()->getLinkCache();
        $this->stats = $statsdDataFactory ?? MediaWikiServices::getInstance()->getStatsdDataFactory();
        $this->logger = $logger ?? LoggerFactory::getInstance( 'runJobs' );
    }
 
    public function run( array $options ) {
        $type = $options['type'] ?? false;
        $maxJobs = $options['maxJobs'] ?? false;
        $maxTime = $options['maxTime'] ?? false;
        $throttle = $options['throttle'] ?? true;
 
        $jobClasses = $this->options->get( 'JobClasses' );
        $profilerLimits = $this->options->get( 'TrxProfilerLimits' );
 
        $response = [ 'jobs' => [], 'reached' => 'none-ready' ];
 
        if ( $type !== false && !isset( $jobClasses[$type] ) ) {
            // Invalid job type specified
            $response['reached'] = 'none-possible';
            return $response;
        }
 
        if ( $this->readOnlyMode->isReadOnly() ) {
            // Any jobs popped off the queue might fail to run and thus might end up lost
            $response['reached'] = 'read-only';
            return $response;
        }
 
        list( , $maxLag ) = $this->lbFactory->getMainLB()->getMaxLag();
        if ( $maxLag >= self::MAX_ALLOWED_LAG ) {
            // DB lag is already too high; caller can immediately try other wikis if applicable
            $response['reached'] = 'replica-lag-limit';
            return $response;
        }
 
        // Narrow DB query expectations for this HTTP request
        $this->lbFactory->getTransactionProfiler()
            ->setExpectations( $profilerLimits['JobRunner'], __METHOD__ );
 
        // Error out if an explicit DB transaction round is somehow active
        if ( $this->lbFactory->hasTransactionRound() ) {
            throw new LogicException( __METHOD__ . ' called with an active transaction round.' );
        }
 
        // Some jobs types should not run until a certain timestamp
        $backoffs = []; // map of (type => UNIX expiry)
        $backoffDeltas = []; // map of (type => seconds)
        $wait = 'wait'; // block to read backoffs the first time
 
        $loopStartTime = microtime( true );
        $jobsPopped = 0;
        $timeMsTotal = 0;
        $lastSyncTime = 1; // initialize "last sync check timestamp" to "ages ago"
        // Keep popping and running jobs until there are no more...
        do {
            // Sync the persistent backoffs with concurrent runners
            $backoffs = $this->syncBackoffDeltas( $backoffs, $backoffDeltas, $wait );
            $backoffKeys = $throttle ? array_keys( $backoffs ) : [];
            $wait = 'nowait'; // less important now
 
            if ( $type === false ) {
                // Treat the default job type queues as a single queue and pop off a job
                $job = $this->jobQueueGroup
                    ->pop( JobQueueGroup::TYPE_DEFAULT, JobQueueGroup::USE_CACHE, $backoffKeys );
            } else {
                // Pop off a job from the specified job type queue unless the execution of
                // that type of job is currently rate-limited by the back-off list
                $job = in_array( $type, $backoffKeys ) ? false : $this->jobQueueGroup->pop( $type );
            }
 
            if ( $job ) {
                ++$jobsPopped;
                $jType = $job->getType();
 
                // Back off of certain jobs for a while (for throttling and for errors)
                $ttw = $this->getBackoffTimeToWait( $job );
                if ( $ttw > 0 ) {
                    // Always add the delta for other runners in case the time running the
                    // job negated the backoff for each individually but not collectively.
                    $backoffDeltas[$jType] = ( $backoffDeltas[$jType] ?? 0 ) + $ttw;
                    $backoffs = $this->syncBackoffDeltas( $backoffs, $backoffDeltas, $wait );
                }
 
                $info = $this->executeJob( $job );
 
                // Mark completed or "one shot only" jobs as resolved
                if ( $info['status'] !== false || !$job->allowRetries() ) {
                    $this->jobQueueGroup->ack( $job );
                }
 
                // Back off of certain jobs for a while (for throttling and for errors)
                if ( $info['status'] === false && mt_rand( 0, 49 ) == 0 ) {
                    $ttw = max( $ttw, $this->getErrorBackoffTTL( $info['caught'] ) );
                    $backoffDeltas[$jType] = ( $backoffDeltas[$jType] ?? 0 ) + $ttw;
                }
 
                $response['jobs'][] = [
                    'type'   => $jType,
                    'status' => ( $info['status'] === false ) ? 'failed' : 'ok',
                    'error'  => $info['error'],
                    'time'   => $info['timeMs']
                ];
                $timeMsTotal += $info['timeMs'];
 
                // Break out if we hit the job count or wall time limits
                if ( $maxJobs && $jobsPopped >= $maxJobs ) {
                    $response['reached'] = 'job-limit';
                    break;
                } elseif ( $maxTime && ( microtime( true ) - $loopStartTime ) > $maxTime ) {
                    $response['reached'] = 'time-limit';
                    break;
                }
 
                // Don't let any of the main DB replica DBs get backed up.
                // This only waits for so long before exiting and letting
                // other wikis in the farm (on different masters) get a chance.
                $timePassed = microtime( true ) - $lastSyncTime;
                if ( $timePassed >= self::LAG_CHECK_PERIOD || $timePassed < 0 ) {
                    $opts = [ 'ifWritesSince' => $lastSyncTime, 'timeout' => self::SYNC_TIMEOUT ];
                    if ( !$this->lbFactory->waitForReplication( $opts ) ) {
                        $response['reached'] = 'replica-lag-limit';
                        break;
                    }
                    $lastSyncTime = microtime( true );
                }
 
                // Abort if nearing OOM to avoid erroring out in the middle of a job
                if ( !$this->checkMemoryOK() ) {
                    $response['reached'] = 'memory-limit';
                    break;
                }
            }
        } while ( $job );
 
        // Sync the persistent backoffs for the next runJobs.php pass
        if ( $backoffDeltas ) {
            $this->syncBackoffDeltas( $backoffs, $backoffDeltas, 'wait' );
        }
 
        $response['backoffs'] = $backoffs;
        $response['elapsed'] = $timeMsTotal;
 
        return $response;
    }
 
    public function executeJob( RunnableJob $job ) {
        $oldRequestId = WebRequest::getRequestId();
        // Temporarily inherit the original ID of the web request that spawned this job
        WebRequest::overrideRequestId( $job->getRequestId() );
        // Use an appropriate timeout to balance lag avoidance and job progress
        $oldTimeout = $this->lbFactory->setDefaultReplicationWaitTimeout( self::SYNC_TIMEOUT );
        try {
            return $this->doExecuteJob( $job );
        } finally {
            $this->lbFactory->setDefaultReplicationWaitTimeout( $oldTimeout );
            WebRequest::overrideRequestId( $oldRequestId );
        }
    }
 
    private function doExecuteJob( RunnableJob $job ) {
        $jType = $job->getType();
        $msg = $job->toString() . " STARTING";
        $this->logger->debug( $msg, [ 'job_type' => $job->getType() ] );
        $this->debugCallback( $msg );
 
        // Clear out title cache data from prior snapshots
        // (e.g. from before JobRunner was invoked in this process)
        $this->linkCache->clear();
 
        // Run the job...
        $caught = [];
        $rssStart = $this->getMaxRssKb();
        $jobStartTime = microtime( true );
        try {
            $fnameTrxOwner = get_class( $job ) . '::run'; // give run() outer scope
            // Flush any pending changes left over from an implicit transaction round
            if ( $job->hasExecutionFlag( $job::JOB_NO_EXPLICIT_TRX_ROUND ) ) {
                $this->lbFactory->commitPrimaryChanges( $fnameTrxOwner ); // new implicit round
            } else {
                $this->lbFactory->beginPrimaryChanges( $fnameTrxOwner ); // new explicit round
            }
            // Clear any stale REPEATABLE-READ snapshots from replica DB connections
            $this->lbFactory->flushReplicaSnapshots( $fnameTrxOwner );
            $status = $job->run();
            $error = $job->getLastError();
            // Commit all pending changes from this job
            $this->commitPrimaryChanges( $job, $fnameTrxOwner );
            // Run any deferred update tasks; doUpdates() manages transactions itself
            DeferredUpdates::doUpdates();
        } catch ( Throwable $e ) {
            MWExceptionHandler::rollbackPrimaryChangesAndLog( $e );
            $status = false;
            $error = get_class( $e ) . ': ' . $e->getMessage();
            $caught[] = get_class( $e );
        }
        // Always attempt to call teardown(), even if Job throws exception
        try {
            $job->tearDown( $status );
        } catch ( Throwable $e ) {
            MWExceptionHandler::logException( $e );
        }
 
        $timeMs = intval( ( microtime( true ) - $jobStartTime ) * 1000 );
        $rssEnd = $this->getMaxRssKb();
 
        // Record how long jobs wait before getting popped
        $readyTs = $job->getReadyTimestamp();
        if ( $readyTs ) {
            $pickupDelay = max( 0, $jobStartTime - $readyTs );
            $this->stats->timing( 'jobqueue.pickup_delay.all', 1000 * $pickupDelay );
            $this->stats->timing( "jobqueue.pickup_delay.$jType", 1000 * $pickupDelay );
        }
        // Record root job age for jobs being run
        $rootTimestamp = $job->getRootJobParams()['rootJobTimestamp'];
        if ( $rootTimestamp ) {
            $age = max( 0, $jobStartTime - wfTimestamp( TS_UNIX, $rootTimestamp ) );
            $this->stats->timing( "jobqueue.pickup_root_age.$jType", 1000 * $age );
        }
        // Track the execution time for jobs
        $this->stats->timing( "jobqueue.run.$jType", $timeMs );
        // Track RSS increases for jobs (in case of memory leaks)
        if ( $rssStart && $rssEnd ) {
            $this->stats->updateCount( "jobqueue.rss_delta.$jType", $rssEnd - $rssStart );
        }
 
        if ( $status === false ) {
            $msg = $job->toString() . " t={job_duration} error={job_error}";
            $this->logger->error( $msg, [
                'job_type' => $job->getType(),
                'job_duration' => $timeMs,
                'job_error' => $error,
            ] );
 
            $msg = $job->toString() . " t=$timeMs error={$error}";
            $this->debugCallback( $msg );
        } else {
            $msg = $job->toString() . " t={job_duration} good";
            $this->logger->info( $msg, [
                'job_type' => $job->getType(),
                'job_duration' => $timeMs,
            ] );
 
            $msg = $job->toString() . " t=$timeMs good";
            $this->debugCallback( $msg );
        }
 
        return [
            'status' => $status,
            'error' => $error,
            'caught' => $caught,
            'timeMs' => $timeMs
        ];
    }
 
    private function getErrorBackoffTTL( array $caught ) {
        return in_array( DBReadOnlyError::class, $caught )
            ? self::READONLY_BACKOFF_TTL
            : self::ERROR_BACKOFF_TTL;
    }
 
    private function getMaxRssKb() {
        $info = getrusage( 0 /* RUSAGE_SELF */ );
        // see https://linux.die.net/man/2/getrusage
        return isset( $info['ru_maxrss'] ) ? (int)$info['ru_maxrss'] : null;
    }
 
    private function getBackoffTimeToWait( RunnableJob $job ) {
        $throttling = $this->options->get( 'JobBackoffThrottling' );
 
        if ( !isset( $throttling[$job->getType()] ) || $job instanceof DuplicateJob ) {
            return 0; // not throttled
        }
 
        $itemsPerSecond = $throttling[$job->getType()];
        if ( $itemsPerSecond <= 0 ) {
            return 0; // not throttled
        }
 
        $seconds = 0;
        if ( $job->workItemCount() > 0 ) {
            $exactSeconds = $job->workItemCount() / $itemsPerSecond;
            // use randomized rounding
            $seconds = floor( $exactSeconds );
            $remainder = $exactSeconds - $seconds;
            $seconds += ( mt_rand() / mt_getrandmax() < $remainder ) ? 1 : 0;
        }
 
        return (int)$seconds;
    }
 
    private function loadBackoffs( array $backoffs, $mode = 'wait' ) {
        $file = wfTempDir() . '/mw-runJobs-backoffs.json';
        if ( is_file( $file ) ) {
            $noblock = ( $mode === 'nowait' ) ? LOCK_NB : 0;
            $handle = fopen( $file, 'rb' );
            if ( !flock( $handle, LOCK_SH | $noblock ) ) {
                fclose( $handle );
                return $backoffs; // don't wait on lock
            }
            $content = stream_get_contents( $handle );
            flock( $handle, LOCK_UN );
            fclose( $handle );
            $ctime = microtime( true );
            $cBackoffs = json_decode( $content, true ) ?: [];
            foreach ( $cBackoffs as $type => $timestamp ) {
                if ( $timestamp < $ctime ) {
                    unset( $cBackoffs[$type] );
                }
            }
        } else {
            $cBackoffs = [];
        }
 
        return $cBackoffs;
    }
 
    private function syncBackoffDeltas( array $backoffs, array &$deltas, $mode = 'wait' ) {
        if ( !$deltas ) {
            return $this->loadBackoffs( $backoffs, $mode );
        }
 
        $noblock = ( $mode === 'nowait' ) ? LOCK_NB : 0;
        $file = wfTempDir() . '/mw-runJobs-backoffs.json';
        $handle = fopen( $file, 'wb+' );
        if ( !flock( $handle, LOCK_EX | $noblock ) ) {
            fclose( $handle );
            return $backoffs; // don't wait on lock
        }
        $ctime = microtime( true );
        $content = stream_get_contents( $handle );
        $cBackoffs = json_decode( $content, true ) ?: [];
        foreach ( $deltas as $type => $seconds ) {
            $cBackoffs[$type] = isset( $cBackoffs[$type] ) && $cBackoffs[$type] >= $ctime
                ? $cBackoffs[$type] + $seconds
                : $ctime + $seconds;
        }
        foreach ( $cBackoffs as $type => $timestamp ) {
            if ( $timestamp < $ctime ) {
                unset( $cBackoffs[$type] );
            }
        }
        ftruncate( $handle, 0 );
        fwrite( $handle, json_encode( $cBackoffs ) );
        flock( $handle, LOCK_UN );
        fclose( $handle );
 
        $deltas = [];
 
        return $cBackoffs;
    }
 
    private function checkMemoryOK() {
        static $maxBytes = null;
        if ( $maxBytes === null ) {
            $m = [];
            if ( preg_match( '!^(\d+)(k|m|g|)$!i', ini_get( 'memory_limit' ), $m ) ) {
                list( , $num, $unit ) = $m;
                $conv = [ 'g' => 1073741824, 'm' => 1048576, 'k' => 1024, '' => 1 ];
                $maxBytes = $num * $conv[strtolower( $unit )];
            } else {
                $maxBytes = 0;
            }
        }
        $usedBytes = memory_get_usage();
        if ( $maxBytes && $usedBytes >= 0.95 * $maxBytes ) {
            $msg = "Detected excessive memory usage ({used_bytes}/{max_bytes}).";
            $this->logger->error( $msg, [
                'used_bytes' => $usedBytes,
                'max_bytes' => $maxBytes,
            ] );
 
            $msg = "Detected excessive memory usage ($usedBytes/$maxBytes).";
            $this->debugCallback( $msg );
 
            return false;
        }
 
        return true;
    }
 
    private function debugCallback( $msg ) {
        if ( $this->debug ) {
            call_user_func_array( $this->debug, [ wfTimestamp( TS_DB ) . " $msg\n" ] );
        }
    }
 
    private function commitPrimaryChanges( RunnableJob $job, $fnameTrxOwner ) {
        $syncThreshold = $this->options->get( 'JobSerialCommitThreshold' );
 
        $time = false;
        $lb = $this->lbFactory->getMainLB();
        if ( $syncThreshold !== false && $lb->hasStreamingReplicaServers() ) {
            // Generally, there is one primary connection to the local DB
            $dbwSerial = $lb->getAnyOpenConnection( $lb->getWriterIndex() );
            // We need natively blocking fast locks
            if ( $dbwSerial && $dbwSerial->namedLocksEnqueue() ) {
                $time = $dbwSerial->pendingWriteQueryDuration( $dbwSerial::ESTIMATE_DB_APPLY );
                if ( $time < $syncThreshold ) {
                    $dbwSerial = false;
                }
            } else {
                $dbwSerial = false;
            }
        } else {
            // There are no replica DBs or writes are all to foreign DB (we don't handle that)
            $dbwSerial = false;
        }
 
        if ( !$dbwSerial ) {
            $this->lbFactory->commitPrimaryChanges(
                $fnameTrxOwner,
                // Abort if any transaction was too big
                [ 'maxWriteDuration' => $this->options->get( 'MaxJobDBWriteDuration' ) ]
            );
 
            return;
        }
 
        $ms = intval( 1000 * $time );
 
        $msg = $job->toString() . " COMMIT ENQUEUED [{job_commit_write_ms}ms of writes]";
        $this->logger->info( $msg, [
            'job_type' => $job->getType(),
            'job_commit_write_ms' => $ms,
        ] );
 
        $msg = $job->toString() . " COMMIT ENQUEUED [{$ms}ms of writes]";
        $this->debugCallback( $msg );
 
        // Wait for an exclusive lock to commit
        if ( !$dbwSerial->lock( 'jobrunner-serial-commit', $fnameTrxOwner, 30 ) ) {
            // This will trigger a rollback in the main loop
            throw new DBError( $dbwSerial, "Timed out waiting on commit queue." );
        }
        $unlocker = new ScopedCallback( static function () use ( $dbwSerial, $fnameTrxOwner ) {
            $dbwSerial->unlock( 'jobrunner-serial-commit', $fnameTrxOwner );
        } );
 
        // Wait for the replica DBs to catch up
        $pos = $lb->getPrimaryPos();
        if ( $pos ) {
            $lb->waitForAll( $pos );
        }
 
        // Actually commit the DB primary changes
        $this->lbFactory->commitPrimaryChanges(
            $fnameTrxOwner,
            // Abort if any transaction was too big
            [ 'maxWriteDuration' => $this->options->get( 'MaxJobDBWriteDuration' ) ]
        );
        ScopedCallback::consume( $unlocker );
    }
}