Examples with TaskDriver org.apache.helix.task.TaskDriver used on opensource projects

Example 31 with TaskDriver

use of org.apache.helix.task.TaskDriver in project helix by apache.

the class GenericHelixController method handleEvent.

/**
 * lock-always: caller always needs to obtain an external lock before call, calls to handleEvent()
 * should be serialized
 * @param event
 */
protected void handleEvent(ClusterEvent event, ClusterDataCache cache) {
    HelixManager manager = event.getAttribute(AttributeName.helixmanager.name());
    if (manager == null) {
        logger.error("No cluster manager in event:" + event.getEventType());
        return;
    }
    if (!manager.isLeader()) {
        logger.error("Cluster manager: " + manager.getInstanceName() + " is not leader for " + manager.getClusterName() + ". Pipeline will not be invoked");
        return;
    }
    // will be excuting in un-paused mode. Which might not be the config in ZK.
    if (_paused) {
        logger.info("Cluster " + manager.getClusterName() + " is paused. Ignoring the event:" + event.getEventType());
        return;
    }
    NotificationContext context = null;
    if (event.getAttribute(AttributeName.changeContext.name()) != null) {
        context = event.getAttribute(AttributeName.changeContext.name());
    }
    if (context != null) {
        if (context.getType() == Type.FINALIZE) {
            stopRebalancingTimers();
            logger.info("Get FINALIZE notification, skip the pipeline. Event :" + event.getEventType());
            return;
        } else {
            // TODO: should be in the initialization of controller.
            if (_cache != null) {
                checkRebalancingTimer(manager, Collections.EMPTY_LIST, _cache.getClusterConfig());
            }
            if (_isMonitoring) {
                event.addAttribute(AttributeName.clusterStatusMonitor.name(), _clusterStatusMonitor);
            }
        }
    }
    // add the cache
    event.addAttribute(AttributeName.ClusterDataCache.name(), cache);
    List<Pipeline> pipelines = cache.isTaskCache() ? _taskRegistry.getPipelinesForEvent(event.getEventType()) : _registry.getPipelinesForEvent(event.getEventType());
    if (pipelines == null || pipelines.size() == 0) {
        logger.info("No " + getPipelineType(cache.isTaskCache()) + " pipeline to run for event:" + event.getEventType());
        return;
    }
    logger.info(String.format("START: Invoking %s controller pipeline for cluster %s event: %s", manager.getClusterName(), getPipelineType(cache.isTaskCache()), event.getEventType()));
    long startTime = System.currentTimeMillis();
    boolean rebalanceFail = false;
    for (Pipeline pipeline : pipelines) {
        try {
            pipeline.handle(event);
            pipeline.finish();
        } catch (Exception e) {
            logger.error("Exception while executing " + getPipelineType(cache.isTaskCache()) + "pipeline: " + pipeline + "for cluster ." + _clusterName + ". Will not continue to next pipeline", e);
            if (e instanceof HelixMetaDataAccessException) {
                rebalanceFail = true;
                // If pipeline failed due to read/write fails to zookeeper, retry the pipeline.
                cache.requireFullRefresh();
                logger.warn("Rebalance pipeline failed due to read failure from zookeeper, cluster: " + _clusterName);
                // only push a retry event when there is no pending event in the corresponding event queue.
                if (isEventQueueEmpty(cache.isTaskCache())) {
                    _continousRebalanceFailureCount++;
                    long delay = getRetryDelay(_continousRebalanceFailureCount);
                    if (delay == 0) {
                        forceRebalance(manager, ClusterEventType.RetryRebalance);
                    } else {
                        _asyncTasksThreadPool.schedule(new RebalanceTask(manager, ClusterEventType.RetryRebalance), delay, TimeUnit.MILLISECONDS);
                    }
                    logger.info("Retry rebalance pipeline with delay " + delay + "ms for cluster: " + _clusterName);
                }
            }
            _clusterStatusMonitor.reportRebalanceFailure();
            break;
        }
    }
    if (!rebalanceFail) {
        _continousRebalanceFailureCount = 0;
    }
    long endTime = System.currentTimeMillis();
    logger.info(String.format("END: Invoking %s controller pipeline for event: %s for cluster %s, took %d ms", getPipelineType(cache.isTaskCache()), event.getEventType(), manager.getClusterName(), (endTime - startTime)));
    if (!cache.isTaskCache()) {
        // report event process durations
        NotificationContext notificationContext = event.getAttribute(AttributeName.changeContext.name());
        long enqueueTime = event.getCreationTime();
        long zkCallbackTime;
        StringBuilder sb = new StringBuilder();
        if (notificationContext != null) {
            zkCallbackTime = notificationContext.getCreationTime();
            if (_isMonitoring) {
                _clusterStatusMonitor.updateClusterEventDuration(ClusterEventMonitor.PhaseName.Callback.name(), enqueueTime - zkCallbackTime);
            }
            sb.append(String.format("Callback time for event: " + event.getEventType() + " took: " + (enqueueTime - zkCallbackTime) + " ms\n"));
        }
        if (_isMonitoring) {
            _clusterStatusMonitor.updateClusterEventDuration(ClusterEventMonitor.PhaseName.InQueue.name(), startTime - enqueueTime);
            _clusterStatusMonitor.updateClusterEventDuration(ClusterEventMonitor.PhaseName.TotalProcessed.name(), endTime - startTime);
        }
        sb.append(String.format("InQueue time for event: " + event.getEventType() + " took: " + (startTime - enqueueTime) + " ms\n"));
        sb.append(String.format("TotalProcessed time for event: " + event.getEventType() + " took: " + (endTime - startTime) + " ms"));
        logger.info(sb.toString());
    } else if (_isMonitoring) {
        // report workflow status
        TaskDriver driver = new TaskDriver(manager);
        _clusterStatusMonitor.refreshWorkflowsStatus(driver);
        _clusterStatusMonitor.refreshJobsStatus(driver);
    }
    // If event handling happens before controller deactivate, the process may write unnecessary
    // MBeans to monitoring after the monitor is disabled.
    // So reset ClusterStatusMonitor according to it's status after all event handling.
    // TODO remove this once clusterStatusMonitor blocks any MBean register on isMonitoring = false.
    resetClusterStatusMonitor();
}