xyuanlu commented on a change in pull request #1532: URL: https://github.com/apache/helix/pull/1532#discussion_r546993916
########## File path: helix-core/src/main/java/org/apache/helix/controller/stages/PerReplicaThrottleStage.java ########## @@ -0,0 +1,982 @@ +package org.apache.helix.controller.stages; + +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ + +import java.util.ArrayList; +import java.util.Collections; +import java.util.Comparator; +import java.util.HashMap; +import java.util.HashSet; +import java.util.LinkedHashMap; +import java.util.List; +import java.util.Map; +import java.util.Set; + +import org.apache.helix.HelixDefinedState; +import org.apache.helix.HelixException; +import org.apache.helix.HelixManager; +import org.apache.helix.api.config.StateTransitionThrottleConfig; +import org.apache.helix.controller.LogUtil; +import org.apache.helix.controller.common.PartitionStateMap; +import org.apache.helix.controller.common.ResourcesStateMap; +import org.apache.helix.controller.dataproviders.ResourceControllerDataProvider; +import org.apache.helix.controller.pipeline.AbstractBaseStage; +import org.apache.helix.controller.pipeline.StageException; +import org.apache.helix.model.BuiltInStateModelDefinitions; +import org.apache.helix.model.ClusterConfig; +import org.apache.helix.model.IdealState; +import org.apache.helix.model.MaintenanceSignal; +import org.apache.helix.model.Message; +import org.apache.helix.model.Partition; +import org.apache.helix.model.Resource; +import org.apache.helix.model.StateModelDefinition; +import org.apache.helix.monitoring.mbeans.ClusterStatusMonitor; +import org.apache.helix.monitoring.mbeans.ResourceMonitor; +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + + +public class PerReplicaThrottleStage extends AbstractBaseStage { + private static final Logger logger = + LoggerFactory.getLogger(PerReplicaThrottleStage.class.getName()); + + private boolean isEmitThrottledMsg = false; + + public PerReplicaThrottleStage() { + this(false); + } + + protected PerReplicaThrottleStage(boolean enableEmitThrottledMsg) { + isEmitThrottledMsg = enableEmitThrottledMsg; + } + + @Override + public void process(ClusterEvent event) throws Exception { + _eventId = event.getEventId(); + + CurrentStateOutput currentStateOutput = event.getAttribute(AttributeName.CURRENT_STATE.name()); + + MessageOutput selectedMessages = event.getAttribute(AttributeName.MESSAGES_SELECTED.name()); + LogUtil.logDebug(logger, _eventId, String.format("selectedMessages is: %s", selectedMessages)); + + Map<String, Resource> resourceToRebalance = + event.getAttribute(AttributeName.RESOURCES_TO_REBALANCE.name()); + ResourceControllerDataProvider cache = + event.getAttribute(AttributeName.ControllerDataProvider.name()); + + if (currentStateOutput == null || selectedMessages == null || resourceToRebalance == null + || cache == null) { + throw new StageException(String.format("Missing attributes in event: %s. " + + "Requires CURRENT_STATE (%s) |BEST_POSSIBLE_STATE (%s) |RESOURCES (%s) |DataCache (%s)", + event, currentStateOutput, selectedMessages, resourceToRebalance, cache)); + } + + ResourcesStateMap retracedResourceStateMap = new ResourcesStateMap(); + List<Message> throttledRecoveryMsg = new ArrayList<>(); + List<Message> throttledLoadMsg = new ArrayList<>(); + MessageOutput output = + compute(event, resourceToRebalance, currentStateOutput, selectedMessages, retracedResourceStateMap, throttledRecoveryMsg, throttledLoadMsg); + + if (logger.isDebugEnabled()) { + LogUtil.logDebug(logger, _eventId, String.format("output is")); + for (String resource : resourceToRebalance.keySet()) { + if (output.getResourceMessages(resource) != null) { + LogUtil.logDebug(logger, _eventId, String.format("resource: %s", resource)); + Map<Partition, List<Message>> partitionListMap = output.getResourceMessages(resource); + for (Partition partition : partitionListMap.keySet()) { + for (Message msg : partitionListMap.get(partition)) { + LogUtil.logDebug(logger, _eventId, String + .format("\tresource: %s, partition: %s, msg: %s", resource, partition, msg)); + } + } + } + } + } + event.addAttribute(AttributeName.PER_REPLICA_OUTPUT_MESSAGES.name(), output); + LogUtil.logDebug(logger,_eventId, String.format("retraceResourceStateMap is: %s", retracedResourceStateMap)); + event.addAttribute(AttributeName.PER_REPLICA_RETRACED_STATES.name(), retracedResourceStateMap); + + if (isEmitThrottledMsg) { + event.addAttribute(AttributeName.PER_REPLICA_THROTTLED_RECOVERY_MESSAGES.name(), throttledRecoveryMsg); + event.addAttribute(AttributeName.PER_REPLICA_THROTTLED_LOAD_MESSAGES.name(), throttledLoadMsg); + } + + // Make sure no instance has more replicas/partitions assigned than maxPartitionPerInstance. If + // it does, pause the rebalance and put the cluster on maintenance mode + int maxPartitionPerInstance = cache.getClusterConfig().getMaxPartitionsPerInstance(); + if (maxPartitionPerInstance > 0) { + validateMaxPartitionsPerInstance(retracedResourceStateMap, maxPartitionPerInstance, cache, event); + } + } + + /** + * Go through every instance in the assignment and check that each instance does NOT have more + * replicas for partitions assigned to it than maxPartitionsPerInstance. If the assignment + * violates this, put the cluster on maintenance mode. + * @param retracedResourceStateMap + * @param maxPartitionPerInstance + */ + private void validateMaxPartitionsPerInstance(ResourcesStateMap retracedResourceStateMap, + int maxPartitionPerInstance, ResourceControllerDataProvider cache, ClusterEvent event) { + Map<String, PartitionStateMap> resourceStatesMap = retracedResourceStateMap.getResourceStatesMap(); + Map<String, Integer> instancePartitionCounts = new HashMap<>(); + + for (String resource : resourceStatesMap.keySet()) { + IdealState idealState = cache.getIdealState(resource); + if (idealState != null + && idealState.getStateModelDefRef().equals(BuiltInStateModelDefinitions.Task.name())) { + // Ignore task here. Task has its own throttling logic + continue; + } + + PartitionStateMap partitionStateMap = resourceStatesMap.get(resource); + Map<Partition, Map<String, String>> stateMaps = partitionStateMap.getStateMap(); + for (Partition p : stateMaps.keySet()) { + Map<String, String> stateMap = stateMaps.get(p); + for (String instance : stateMap.keySet()) { + // If this replica is in DROPPED state, do not count it in the partition count since it is + // to be dropped + String state = stateMap.get(instance); + if (state.equals(HelixDefinedState.DROPPED.name())) { + continue; + } + if (!instancePartitionCounts.containsKey(instance)) { + instancePartitionCounts.put(instance, 0); + } + int partitionCount = instancePartitionCounts.get(instance); + // Number of replicas (from different partitions) held in this instance + partitionCount++; + if (partitionCount > maxPartitionPerInstance) { + HelixManager manager = event.getAttribute(AttributeName.helixmanager.name()); + String errMsg = String.format( + "Problem: according to this assignment, instance %s contains more " + + "replicas/partitions than the maximum number allowed (%d). Pipeline will " + + "stop the rebalance and put the cluster %s into maintenance mode", + instance, maxPartitionPerInstance, cache.getClusterName()); + if (manager != null) { + if (manager.getHelixDataAccessor() + .getProperty(manager.getHelixDataAccessor().keyBuilder().maintenance()) == null) { + manager.getClusterManagmentTool().autoEnableMaintenanceMode( + manager.getClusterName(), true, errMsg, + MaintenanceSignal.AutoTriggerReason.MAX_PARTITION_PER_INSTANCE_EXCEEDED); + } + LogUtil.logWarn(logger, _eventId, errMsg); + } else { + LogUtil.logError(logger, _eventId, + "HelixManager is not set/null! Failed to pause this cluster/enable maintenance" + + " mode due to an instance being assigned more replicas/partitions than " + + "the limit."); + } + + ClusterStatusMonitor clusterStatusMonitor = + event.getAttribute(AttributeName.clusterStatusMonitor.name()); + if (clusterStatusMonitor != null) { + clusterStatusMonitor.setResourceRebalanceStates(Collections.singletonList(resource), + ResourceMonitor.RebalanceStatus.PER_REPLICA_STATE_CAL_FAILED); + } + throw new HelixException(errMsg); + } + instancePartitionCounts.put(instance, partitionCount); + } + } + } + } + + /** + * Go through each resource, and based on messageSelected and currentState, compute + * messageOutput while maintaining throttling constraints (for example, ensure that the number + * of possible pending state transitions does NOT go over the set threshold). + * @param event + * @param resourceMap + * @param currentStateOutput + * @param selectedMessage + * @param retracedResourceStateMap out + * @return + */ + private MessageOutput compute(ClusterEvent event, Map<String, Resource> resourceMap, + CurrentStateOutput currentStateOutput, MessageOutput selectedMessage, + ResourcesStateMap retracedResourceStateMap, + List<Message> throttledRecoveryMsg, List<Message> throttledLoadMsg) { + MessageOutput output = new MessageOutput(); + + ResourceControllerDataProvider dataCache = + event.getAttribute(AttributeName.ControllerDataProvider.name()); + + StateTransitionThrottleController throttleController = + new StateTransitionThrottleController(resourceMap.keySet(), dataCache.getClusterConfig(), + dataCache.getLiveInstances().keySet()); + + // Resource level prioritization based on the numerical (sortable) priority field. + // If the resource priority field is null/not set, the resource will be treated as lowest + // priority. + List<ResourcePriority> prioritizedResourceList = new ArrayList<>(); + for (String resourceName : resourceMap.keySet()) { + prioritizedResourceList.add(new ResourcePriority(resourceName, Integer.MIN_VALUE)); + } + // If resourcePriorityField is null at the cluster level, all resources will be considered equal + // in priority by keeping all priorities at MIN_VALUE + String priorityField = dataCache.getClusterConfig().getResourcePriorityField(); + if (priorityField != null) { + for (ResourcePriority resourcePriority : prioritizedResourceList) { + String resourceName = resourcePriority.getResourceName(); + + // Will take the priority from ResourceConfig first + // If ResourceConfig does not exist or does not have this field. + // Try to load it from the resource's IdealState. Otherwise, keep it at the lowest priority + if (dataCache.getResourceConfig(resourceName) != null + && dataCache.getResourceConfig(resourceName).getSimpleConfig(priorityField) != null) { + resourcePriority.setPriority( + dataCache.getResourceConfig(resourceName).getSimpleConfig(priorityField)); + } else if (dataCache.getIdealState(resourceName) != null + && dataCache.getIdealState(resourceName).getRecord().getSimpleField(priorityField) + != null) { + resourcePriority.setPriority( + dataCache.getIdealState(resourceName).getRecord().getSimpleField(priorityField)); + } + } + prioritizedResourceList.sort(new ResourcePriorityComparator()); + } + + ClusterStatusMonitor clusterStatusMonitor = + event.getAttribute(AttributeName.clusterStatusMonitor.name()); + List<String> failedResources = new ArrayList<>(); + + // Priority is applied in assignment computation because higher priority by looping in order of + // decreasing priority + for (ResourcePriority resourcePriority : prioritizedResourceList) { + String resourceName = resourcePriority.getResourceName(); + + BestPossibleStateOutput bestPossibleStateOutput = + event.getAttribute(AttributeName.BEST_POSSIBLE_STATE.name()); + if (!bestPossibleStateOutput.containsResource(resourceName)) { + LogUtil.logInfo(logger, _eventId, String.format( + "Skip calculating per replica state for resource %s because the best possible state is not available.", + resourceName)); + continue; + } + + Resource resource = resourceMap.get(resourceName); + IdealState idealState = dataCache.getIdealState(resourceName); + if (idealState == null) { + // If IdealState is null, use an empty one + LogUtil.logInfo(logger, _eventId, String + .format("IdealState for resource %s does not exist; resource may not exist anymore", + resourceName)); + idealState = new IdealState(resourceName); + idealState.setStateModelDefRef(resource.getStateModelDefRef()); + } + + Map<Partition, Map<String, String>> retracedPartitionsState = new HashMap<>(); + try { + throttlePerReplicaMessages(idealState, clusterStatusMonitor, currentStateOutput, + selectedMessage.getResourceMessages(resourceName), resourceMap.get(resourceName), + bestPossibleStateOutput, dataCache, throttleController, retracedPartitionsState, + throttledRecoveryMsg, throttledLoadMsg, output); + retracedResourceStateMap.setState(resourceName, retracedPartitionsState); + } catch (HelixException ex) { + LogUtil.logInfo(logger, _eventId, + "Failed to calculate per replica partition states for resource " + resourceName, ex); + failedResources.add(resourceName); + } + } + + if (clusterStatusMonitor != null) { + clusterStatusMonitor.setResourceRebalanceStates(failedResources, + ResourceMonitor.RebalanceStatus.PER_REPLICA_STATE_CAL_FAILED); + clusterStatusMonitor.setResourceRebalanceStates(output.resourceSet(), + ResourceMonitor.RebalanceStatus.NORMAL); + } + + return output; + } + + /* + * Apply per-replica throttling logic and filter out excessive recovery and load messages for a + * given resource. + * Reconstruct retrace partition states for a resource based on pending and targeted messages + * Return messages for partitions of a resource. + * Out param retracedPartitionsCurrentState + * Out param output + */ + private void throttlePerReplicaMessages(IdealState idealState, + ClusterStatusMonitor clusterStatusMonitor, + CurrentStateOutput currentStateOutput, Map<Partition, List<Message>> selectedResourceMessages, + Resource resource, BestPossibleStateOutput bestPossibleStateOutput, + ResourceControllerDataProvider cache, StateTransitionThrottleController throttleController, + Map<Partition, Map<String, String>> retracedPartitionsStateMap, + List<Message> throttledRecoveryMsgOut, List<Message> throttledLoadMessageOut, + MessageOutput output) { + String resourceName = resource.getResourceName(); + LogUtil.logInfo(logger, _eventId, String.format("Processing resource: %s", resourceName)); + + if (!throttleController.isThrottleEnabled() || !IdealState.RebalanceMode.FULL_AUTO + .equals(idealState.getRebalanceMode())) { + retracedPartitionsStateMap.putAll(bestPossibleStateOutput.getPartitionStateMap(resourceName).getStateMap()); + for (Partition partition : selectedResourceMessages.keySet()) { + output.addMessages(resourceName, partition, selectedResourceMessages.get(partition)); + } + return; + } + Map<String, List<String>> preferenceLists = + bestPossibleStateOutput.getPreferenceLists(resourceName); + + Set<Partition> partitionsWithErrorStateReplica = new HashSet<>(); + Set<Partition> partitionsNeedRecovery = new HashSet<>(); + + // Step 1: charge existing pending messages and update retraced state map. + chargePendingMessages(resource, throttleController, currentStateOutput, bestPossibleStateOutput, + idealState, cache, partitionsNeedRecovery, partitionsWithErrorStateReplica, + retracedPartitionsStateMap); + + // Step 2: classify all the messages into recovery message list and load message list + List<Message> recoveryMessages = new ArrayList<>(); + List<Message> loadMessages = new ArrayList<>(); + Map<Message, Partition> messagePartitionMap = new HashMap<>(); // todo: Message may need a hashcode() + classifyMessages(resource, currentStateOutput, bestPossibleStateOutput, idealState, cache, + selectedResourceMessages, recoveryMessages, loadMessages, messagePartitionMap); + + // Step 3: sorts recovery message list and applies throttling + Set<Message> throttledRecoveryMessages = new HashSet<>(); + + Map<Partition, Map<String, String>> bestPossibleMap = + bestPossibleStateOutput.getPartitionStateMap(resourceName).getStateMap(); + Map<Partition, Map<String, String>> currentStateMap = + currentStateOutput.getCurrentStateMap(resourceName); + String stateModelDefName = idealState.getStateModelDefRef(); + StateModelDefinition stateModelDef = cache.getStateModelDef(stateModelDefName); + LogUtil.logDebug(logger, _eventId, + String.format("applying recovery rebalance with resource %s", resourceName)); + applyThrottling(resource, throttleController, currentStateMap, bestPossibleMap, idealState, + cache, false, recoveryMessages, messagePartitionMap, + throttledRecoveryMessages, StateTransitionThrottleConfig.RebalanceType.RECOVERY_BALANCE); + + // Step 4: sorts load message list and applies throttling + + // TODO: consider simplifying the following logic. + // calculate error-on-recovery downward flag + // If the threshold (ErrorOrRecovery) is set, then use it, if not, then check if the old + // threshold (Error) is set. If the old threshold is set, use it. If not, use the default value + // for the new one. This is for backward-compatibility + int threshold = 1; // Default threshold for ErrorOrRecoveryPartitionThresholdForLoadBalance + int partitionCount = partitionsWithErrorStateReplica.size(); + ClusterConfig clusterConfig = cache.getClusterConfig(); + if (clusterConfig.getErrorOrRecoveryPartitionThresholdForLoadBalance() != -1) { + // ErrorOrRecovery is set + threshold = clusterConfig.getErrorOrRecoveryPartitionThresholdForLoadBalance(); + partitionCount += partitionsNeedRecovery.size(); // Only add this count when the threshold is set + } else { + if (clusterConfig.getErrorPartitionThresholdForLoadBalance() != 0) { + // 0 is the default value so the old threshold has been set + threshold = clusterConfig.getErrorPartitionThresholdForLoadBalance(); + } + } + + // Perform regular load balance only if the number of partitions in recovery and in error is + // less than the threshold. Otherwise, only allow downward-transition load balance + boolean onlyDownwardLoadBalance = partitionCount > threshold; + Set<Message> throttledLoadMessages = new HashSet<>(); + LogUtil.logDebug(logger, _eventId, + String.format("applying load rebalance with resource %s, onlyDownwardLoadBalance %s", + resourceName, onlyDownwardLoadBalance)); + applyThrottling(resource, throttleController, currentStateMap, bestPossibleMap, idealState, + cache, onlyDownwardLoadBalance, loadMessages, messagePartitionMap, throttledLoadMessages, + StateTransitionThrottleConfig.RebalanceType.LOAD_BALANCE); + + LogUtil.logDebug(logger, _eventId, + String.format("resource %s, throttled recovery message: %s", resourceName,throttledRecoveryMessages)); + LogUtil.logDebug(logger, _eventId, + String.format("resource %s, throttled load messages: %s", resourceName,throttledLoadMessages)); + + throttledRecoveryMsgOut.addAll(throttledRecoveryMessages); + throttledLoadMessageOut.addAll(throttledLoadMessages); + + // Step 5: construct output + Map<Partition, List<Message>> out = new HashMap<>(); + for (Partition partition : resource.getPartitions()) { + List<Message> partitionMessages = selectedResourceMessages.get(partition); + if (partitionMessages == null) { + continue; + } + List<Message> finalPartitionMessages = new ArrayList<>(); + for (Message message: partitionMessages) { + if (throttledRecoveryMessages.contains(message)) { + continue; + } + if (throttledLoadMessages.contains(message)) { + continue; + } + finalPartitionMessages.add(message); + } + out.put(partition, finalPartitionMessages); + output.addMessages(resourceName, partition, finalPartitionMessages); + } + + // Step 6: constructs all retraced partition state map for the resource; + constructRetracedPartitionStateMap(resource, retracedPartitionsStateMap, out); + + // Step 7: emit metrics + if (clusterStatusMonitor != null) { + clusterStatusMonitor.updatePerReplicaRebalancerStats(resourceName, recoveryMessages.size(), + loadMessages.size(), throttledRecoveryMsgOut.size(), + throttledLoadMessageOut.size()); + } + } + + private void constructRetracedPartitionStateMap(Resource resource, + Map<Partition, Map<String, String>> retracedPartitionsStateMap, + Map<Partition, List<Message>> out) { + for (Partition partition : resource.getPartitions()) { + List<Message> partitionMessages = out.get(partition); + if (partitionMessages == null) { + continue; + } + for (Message message : partitionMessages) { + if (!Message.MessageType.STATE_TRANSITION.name().equals(message.getMsgType())) { + // todo: log? + // ignore cancellation message etc. + continue; + } + String toState = message.getToState(); + // toIntance may not be in the retracedStateMap as so far it is current state based. + // new instance in best possible not in currentstate would not be in retracedStateMap yet. + String toInstance = message.getTgtName(); + Map<String, String> retracedStateMap = retracedPartitionsStateMap.get(partition); + retracedStateMap.put(toInstance, toState); Review comment: nit: retracedStateMap.put(message.getTgtName(), message.getToState()); might be cleaner. ########## File path: helix-core/src/main/java/org/apache/helix/controller/stages/PerReplicaThrottleStage.java ########## @@ -0,0 +1,982 @@ +package org.apache.helix.controller.stages; + +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ + +import java.util.ArrayList; +import java.util.Collections; +import java.util.Comparator; +import java.util.HashMap; +import java.util.HashSet; +import java.util.LinkedHashMap; +import java.util.List; +import java.util.Map; +import java.util.Set; + +import org.apache.helix.HelixDefinedState; +import org.apache.helix.HelixException; +import org.apache.helix.HelixManager; +import org.apache.helix.api.config.StateTransitionThrottleConfig; +import org.apache.helix.controller.LogUtil; +import org.apache.helix.controller.common.PartitionStateMap; +import org.apache.helix.controller.common.ResourcesStateMap; +import org.apache.helix.controller.dataproviders.ResourceControllerDataProvider; +import org.apache.helix.controller.pipeline.AbstractBaseStage; +import org.apache.helix.controller.pipeline.StageException; +import org.apache.helix.model.BuiltInStateModelDefinitions; +import org.apache.helix.model.ClusterConfig; +import org.apache.helix.model.IdealState; +import org.apache.helix.model.MaintenanceSignal; +import org.apache.helix.model.Message; +import org.apache.helix.model.Partition; +import org.apache.helix.model.Resource; +import org.apache.helix.model.StateModelDefinition; +import org.apache.helix.monitoring.mbeans.ClusterStatusMonitor; +import org.apache.helix.monitoring.mbeans.ResourceMonitor; +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + + +public class PerReplicaThrottleStage extends AbstractBaseStage { + private static final Logger logger = + LoggerFactory.getLogger(PerReplicaThrottleStage.class.getName()); + + private boolean isEmitThrottledMsg = false; + + public PerReplicaThrottleStage() { + this(false); + } + + protected PerReplicaThrottleStage(boolean enableEmitThrottledMsg) { + isEmitThrottledMsg = enableEmitThrottledMsg; + } + + @Override + public void process(ClusterEvent event) throws Exception { + _eventId = event.getEventId(); + + CurrentStateOutput currentStateOutput = event.getAttribute(AttributeName.CURRENT_STATE.name()); + + MessageOutput selectedMessages = event.getAttribute(AttributeName.MESSAGES_SELECTED.name()); + LogUtil.logDebug(logger, _eventId, String.format("selectedMessages is: %s", selectedMessages)); + + Map<String, Resource> resourceToRebalance = + event.getAttribute(AttributeName.RESOURCES_TO_REBALANCE.name()); + ResourceControllerDataProvider cache = + event.getAttribute(AttributeName.ControllerDataProvider.name()); + + if (currentStateOutput == null || selectedMessages == null || resourceToRebalance == null + || cache == null) { + throw new StageException(String.format("Missing attributes in event: %s. " + + "Requires CURRENT_STATE (%s) |BEST_POSSIBLE_STATE (%s) |RESOURCES (%s) |DataCache (%s)", + event, currentStateOutput, selectedMessages, resourceToRebalance, cache)); + } + + ResourcesStateMap retracedResourceStateMap = new ResourcesStateMap(); + List<Message> throttledRecoveryMsg = new ArrayList<>(); + List<Message> throttledLoadMsg = new ArrayList<>(); + MessageOutput output = + compute(event, resourceToRebalance, currentStateOutput, selectedMessages, retracedResourceStateMap, throttledRecoveryMsg, throttledLoadMsg); + + if (logger.isDebugEnabled()) { + LogUtil.logDebug(logger, _eventId, String.format("output is")); + for (String resource : resourceToRebalance.keySet()) { + if (output.getResourceMessages(resource) != null) { + LogUtil.logDebug(logger, _eventId, String.format("resource: %s", resource)); + Map<Partition, List<Message>> partitionListMap = output.getResourceMessages(resource); + for (Partition partition : partitionListMap.keySet()) { + for (Message msg : partitionListMap.get(partition)) { + LogUtil.logDebug(logger, _eventId, String + .format("\tresource: %s, partition: %s, msg: %s", resource, partition, msg)); + } + } + } + } + } + event.addAttribute(AttributeName.PER_REPLICA_OUTPUT_MESSAGES.name(), output); + LogUtil.logDebug(logger,_eventId, String.format("retraceResourceStateMap is: %s", retracedResourceStateMap)); + event.addAttribute(AttributeName.PER_REPLICA_RETRACED_STATES.name(), retracedResourceStateMap); + + if (isEmitThrottledMsg) { + event.addAttribute(AttributeName.PER_REPLICA_THROTTLED_RECOVERY_MESSAGES.name(), throttledRecoveryMsg); + event.addAttribute(AttributeName.PER_REPLICA_THROTTLED_LOAD_MESSAGES.name(), throttledLoadMsg); + } + + // Make sure no instance has more replicas/partitions assigned than maxPartitionPerInstance. If + // it does, pause the rebalance and put the cluster on maintenance mode + int maxPartitionPerInstance = cache.getClusterConfig().getMaxPartitionsPerInstance(); + if (maxPartitionPerInstance > 0) { + validateMaxPartitionsPerInstance(retracedResourceStateMap, maxPartitionPerInstance, cache, event); + } + } + + /** + * Go through every instance in the assignment and check that each instance does NOT have more + * replicas for partitions assigned to it than maxPartitionsPerInstance. If the assignment + * violates this, put the cluster on maintenance mode. + * @param retracedResourceStateMap + * @param maxPartitionPerInstance + */ + private void validateMaxPartitionsPerInstance(ResourcesStateMap retracedResourceStateMap, + int maxPartitionPerInstance, ResourceControllerDataProvider cache, ClusterEvent event) { + Map<String, PartitionStateMap> resourceStatesMap = retracedResourceStateMap.getResourceStatesMap(); + Map<String, Integer> instancePartitionCounts = new HashMap<>(); + + for (String resource : resourceStatesMap.keySet()) { + IdealState idealState = cache.getIdealState(resource); + if (idealState != null + && idealState.getStateModelDefRef().equals(BuiltInStateModelDefinitions.Task.name())) { + // Ignore task here. Task has its own throttling logic + continue; + } + + PartitionStateMap partitionStateMap = resourceStatesMap.get(resource); + Map<Partition, Map<String, String>> stateMaps = partitionStateMap.getStateMap(); + for (Partition p : stateMaps.keySet()) { + Map<String, String> stateMap = stateMaps.get(p); + for (String instance : stateMap.keySet()) { + // If this replica is in DROPPED state, do not count it in the partition count since it is + // to be dropped + String state = stateMap.get(instance); + if (state.equals(HelixDefinedState.DROPPED.name())) { + continue; + } + if (!instancePartitionCounts.containsKey(instance)) { + instancePartitionCounts.put(instance, 0); + } + int partitionCount = instancePartitionCounts.get(instance); + // Number of replicas (from different partitions) held in this instance + partitionCount++; + if (partitionCount > maxPartitionPerInstance) { + HelixManager manager = event.getAttribute(AttributeName.helixmanager.name()); + String errMsg = String.format( + "Problem: according to this assignment, instance %s contains more " + + "replicas/partitions than the maximum number allowed (%d). Pipeline will " + + "stop the rebalance and put the cluster %s into maintenance mode", + instance, maxPartitionPerInstance, cache.getClusterName()); + if (manager != null) { + if (manager.getHelixDataAccessor() + .getProperty(manager.getHelixDataAccessor().keyBuilder().maintenance()) == null) { + manager.getClusterManagmentTool().autoEnableMaintenanceMode( + manager.getClusterName(), true, errMsg, + MaintenanceSignal.AutoTriggerReason.MAX_PARTITION_PER_INSTANCE_EXCEEDED); + } + LogUtil.logWarn(logger, _eventId, errMsg); + } else { + LogUtil.logError(logger, _eventId, + "HelixManager is not set/null! Failed to pause this cluster/enable maintenance" + + " mode due to an instance being assigned more replicas/partitions than " + + "the limit."); + } + + ClusterStatusMonitor clusterStatusMonitor = + event.getAttribute(AttributeName.clusterStatusMonitor.name()); + if (clusterStatusMonitor != null) { + clusterStatusMonitor.setResourceRebalanceStates(Collections.singletonList(resource), + ResourceMonitor.RebalanceStatus.PER_REPLICA_STATE_CAL_FAILED); + } + throw new HelixException(errMsg); + } + instancePartitionCounts.put(instance, partitionCount); + } + } + } + } + + /** + * Go through each resource, and based on messageSelected and currentState, compute + * messageOutput while maintaining throttling constraints (for example, ensure that the number + * of possible pending state transitions does NOT go over the set threshold). + * @param event + * @param resourceMap + * @param currentStateOutput + * @param selectedMessage + * @param retracedResourceStateMap out + * @return + */ + private MessageOutput compute(ClusterEvent event, Map<String, Resource> resourceMap, + CurrentStateOutput currentStateOutput, MessageOutput selectedMessage, + ResourcesStateMap retracedResourceStateMap, + List<Message> throttledRecoveryMsg, List<Message> throttledLoadMsg) { + MessageOutput output = new MessageOutput(); + + ResourceControllerDataProvider dataCache = + event.getAttribute(AttributeName.ControllerDataProvider.name()); + + StateTransitionThrottleController throttleController = + new StateTransitionThrottleController(resourceMap.keySet(), dataCache.getClusterConfig(), + dataCache.getLiveInstances().keySet()); + + // Resource level prioritization based on the numerical (sortable) priority field. + // If the resource priority field is null/not set, the resource will be treated as lowest + // priority. + List<ResourcePriority> prioritizedResourceList = new ArrayList<>(); + for (String resourceName : resourceMap.keySet()) { + prioritizedResourceList.add(new ResourcePriority(resourceName, Integer.MIN_VALUE)); + } + // If resourcePriorityField is null at the cluster level, all resources will be considered equal + // in priority by keeping all priorities at MIN_VALUE + String priorityField = dataCache.getClusterConfig().getResourcePriorityField(); + if (priorityField != null) { + for (ResourcePriority resourcePriority : prioritizedResourceList) { + String resourceName = resourcePriority.getResourceName(); + + // Will take the priority from ResourceConfig first + // If ResourceConfig does not exist or does not have this field. + // Try to load it from the resource's IdealState. Otherwise, keep it at the lowest priority + if (dataCache.getResourceConfig(resourceName) != null + && dataCache.getResourceConfig(resourceName).getSimpleConfig(priorityField) != null) { + resourcePriority.setPriority( + dataCache.getResourceConfig(resourceName).getSimpleConfig(priorityField)); + } else if (dataCache.getIdealState(resourceName) != null + && dataCache.getIdealState(resourceName).getRecord().getSimpleField(priorityField) + != null) { + resourcePriority.setPriority( + dataCache.getIdealState(resourceName).getRecord().getSimpleField(priorityField)); + } + } + prioritizedResourceList.sort(new ResourcePriorityComparator()); + } + + ClusterStatusMonitor clusterStatusMonitor = + event.getAttribute(AttributeName.clusterStatusMonitor.name()); + List<String> failedResources = new ArrayList<>(); + + // Priority is applied in assignment computation because higher priority by looping in order of + // decreasing priority + for (ResourcePriority resourcePriority : prioritizedResourceList) { + String resourceName = resourcePriority.getResourceName(); + + BestPossibleStateOutput bestPossibleStateOutput = + event.getAttribute(AttributeName.BEST_POSSIBLE_STATE.name()); + if (!bestPossibleStateOutput.containsResource(resourceName)) { + LogUtil.logInfo(logger, _eventId, String.format( + "Skip calculating per replica state for resource %s because the best possible state is not available.", + resourceName)); + continue; + } + + Resource resource = resourceMap.get(resourceName); + IdealState idealState = dataCache.getIdealState(resourceName); + if (idealState == null) { + // If IdealState is null, use an empty one + LogUtil.logInfo(logger, _eventId, String + .format("IdealState for resource %s does not exist; resource may not exist anymore", + resourceName)); + idealState = new IdealState(resourceName); + idealState.setStateModelDefRef(resource.getStateModelDefRef()); + } + + Map<Partition, Map<String, String>> retracedPartitionsState = new HashMap<>(); + try { + throttlePerReplicaMessages(idealState, clusterStatusMonitor, currentStateOutput, + selectedMessage.getResourceMessages(resourceName), resourceMap.get(resourceName), + bestPossibleStateOutput, dataCache, throttleController, retracedPartitionsState, + throttledRecoveryMsg, throttledLoadMsg, output); + retracedResourceStateMap.setState(resourceName, retracedPartitionsState); + } catch (HelixException ex) { + LogUtil.logInfo(logger, _eventId, + "Failed to calculate per replica partition states for resource " + resourceName, ex); + failedResources.add(resourceName); + } + } + + if (clusterStatusMonitor != null) { + clusterStatusMonitor.setResourceRebalanceStates(failedResources, + ResourceMonitor.RebalanceStatus.PER_REPLICA_STATE_CAL_FAILED); + clusterStatusMonitor.setResourceRebalanceStates(output.resourceSet(), + ResourceMonitor.RebalanceStatus.NORMAL); + } + + return output; + } + + /* + * Apply per-replica throttling logic and filter out excessive recovery and load messages for a + * given resource. + * Reconstruct retrace partition states for a resource based on pending and targeted messages + * Return messages for partitions of a resource. + * Out param retracedPartitionsCurrentState + * Out param output + */ + private void throttlePerReplicaMessages(IdealState idealState, + ClusterStatusMonitor clusterStatusMonitor, + CurrentStateOutput currentStateOutput, Map<Partition, List<Message>> selectedResourceMessages, + Resource resource, BestPossibleStateOutput bestPossibleStateOutput, + ResourceControllerDataProvider cache, StateTransitionThrottleController throttleController, + Map<Partition, Map<String, String>> retracedPartitionsStateMap, + List<Message> throttledRecoveryMsgOut, List<Message> throttledLoadMessageOut, + MessageOutput output) { + String resourceName = resource.getResourceName(); + LogUtil.logInfo(logger, _eventId, String.format("Processing resource: %s", resourceName)); + + if (!throttleController.isThrottleEnabled() || !IdealState.RebalanceMode.FULL_AUTO + .equals(idealState.getRebalanceMode())) { + retracedPartitionsStateMap.putAll(bestPossibleStateOutput.getPartitionStateMap(resourceName).getStateMap()); + for (Partition partition : selectedResourceMessages.keySet()) { + output.addMessages(resourceName, partition, selectedResourceMessages.get(partition)); + } + return; + } + Map<String, List<String>> preferenceLists = + bestPossibleStateOutput.getPreferenceLists(resourceName); + + Set<Partition> partitionsWithErrorStateReplica = new HashSet<>(); + Set<Partition> partitionsNeedRecovery = new HashSet<>(); + + // Step 1: charge existing pending messages and update retraced state map. + chargePendingMessages(resource, throttleController, currentStateOutput, bestPossibleStateOutput, + idealState, cache, partitionsNeedRecovery, partitionsWithErrorStateReplica, + retracedPartitionsStateMap); + + // Step 2: classify all the messages into recovery message list and load message list + List<Message> recoveryMessages = new ArrayList<>(); + List<Message> loadMessages = new ArrayList<>(); + Map<Message, Partition> messagePartitionMap = new HashMap<>(); // todo: Message may need a hashcode() + classifyMessages(resource, currentStateOutput, bestPossibleStateOutput, idealState, cache, + selectedResourceMessages, recoveryMessages, loadMessages, messagePartitionMap); + + // Step 3: sorts recovery message list and applies throttling + Set<Message> throttledRecoveryMessages = new HashSet<>(); + + Map<Partition, Map<String, String>> bestPossibleMap = + bestPossibleStateOutput.getPartitionStateMap(resourceName).getStateMap(); + Map<Partition, Map<String, String>> currentStateMap = + currentStateOutput.getCurrentStateMap(resourceName); + String stateModelDefName = idealState.getStateModelDefRef(); + StateModelDefinition stateModelDef = cache.getStateModelDef(stateModelDefName); + LogUtil.logDebug(logger, _eventId, + String.format("applying recovery rebalance with resource %s", resourceName)); + applyThrottling(resource, throttleController, currentStateMap, bestPossibleMap, idealState, + cache, false, recoveryMessages, messagePartitionMap, + throttledRecoveryMessages, StateTransitionThrottleConfig.RebalanceType.RECOVERY_BALANCE); + + // Step 4: sorts load message list and applies throttling + + // TODO: consider simplifying the following logic. + // calculate error-on-recovery downward flag + // If the threshold (ErrorOrRecovery) is set, then use it, if not, then check if the old + // threshold (Error) is set. If the old threshold is set, use it. If not, use the default value + // for the new one. This is for backward-compatibility + int threshold = 1; // Default threshold for ErrorOrRecoveryPartitionThresholdForLoadBalance + int partitionCount = partitionsWithErrorStateReplica.size(); + ClusterConfig clusterConfig = cache.getClusterConfig(); + if (clusterConfig.getErrorOrRecoveryPartitionThresholdForLoadBalance() != -1) { + // ErrorOrRecovery is set + threshold = clusterConfig.getErrorOrRecoveryPartitionThresholdForLoadBalance(); + partitionCount += partitionsNeedRecovery.size(); // Only add this count when the threshold is set + } else { + if (clusterConfig.getErrorPartitionThresholdForLoadBalance() != 0) { + // 0 is the default value so the old threshold has been set + threshold = clusterConfig.getErrorPartitionThresholdForLoadBalance(); + } + } + + // Perform regular load balance only if the number of partitions in recovery and in error is + // less than the threshold. Otherwise, only allow downward-transition load balance + boolean onlyDownwardLoadBalance = partitionCount > threshold; + Set<Message> throttledLoadMessages = new HashSet<>(); + LogUtil.logDebug(logger, _eventId, + String.format("applying load rebalance with resource %s, onlyDownwardLoadBalance %s", + resourceName, onlyDownwardLoadBalance)); + applyThrottling(resource, throttleController, currentStateMap, bestPossibleMap, idealState, + cache, onlyDownwardLoadBalance, loadMessages, messagePartitionMap, throttledLoadMessages, + StateTransitionThrottleConfig.RebalanceType.LOAD_BALANCE); + + LogUtil.logDebug(logger, _eventId, + String.format("resource %s, throttled recovery message: %s", resourceName,throttledRecoveryMessages)); + LogUtil.logDebug(logger, _eventId, + String.format("resource %s, throttled load messages: %s", resourceName,throttledLoadMessages)); + + throttledRecoveryMsgOut.addAll(throttledRecoveryMessages); + throttledLoadMessageOut.addAll(throttledLoadMessages); + + // Step 5: construct output + Map<Partition, List<Message>> out = new HashMap<>(); + for (Partition partition : resource.getPartitions()) { + List<Message> partitionMessages = selectedResourceMessages.get(partition); + if (partitionMessages == null) { + continue; + } + List<Message> finalPartitionMessages = new ArrayList<>(); + for (Message message: partitionMessages) { + if (throttledRecoveryMessages.contains(message)) { + continue; + } + if (throttledLoadMessages.contains(message)) { + continue; + } + finalPartitionMessages.add(message); + } + out.put(partition, finalPartitionMessages); + output.addMessages(resourceName, partition, finalPartitionMessages); + } + + // Step 6: constructs all retraced partition state map for the resource; + constructRetracedPartitionStateMap(resource, retracedPartitionsStateMap, out); + + // Step 7: emit metrics + if (clusterStatusMonitor != null) { + clusterStatusMonitor.updatePerReplicaRebalancerStats(resourceName, recoveryMessages.size(), + loadMessages.size(), throttledRecoveryMsgOut.size(), + throttledLoadMessageOut.size()); + } + } + + private void constructRetracedPartitionStateMap(Resource resource, + Map<Partition, Map<String, String>> retracedPartitionsStateMap, + Map<Partition, List<Message>> out) { + for (Partition partition : resource.getPartitions()) { + List<Message> partitionMessages = out.get(partition); + if (partitionMessages == null) { + continue; + } + for (Message message : partitionMessages) { + if (!Message.MessageType.STATE_TRANSITION.name().equals(message.getMsgType())) { + // todo: log? + // ignore cancellation message etc. + continue; + } + String toState = message.getToState(); + // toIntance may not be in the retracedStateMap as so far it is current state based. + // new instance in best possible not in currentstate would not be in retracedStateMap yet. + String toInstance = message.getTgtName(); + Map<String, String> retracedStateMap = retracedPartitionsStateMap.get(partition); + retracedStateMap.put(toInstance, toState); Review comment: nit: ``` retracedStateMap.put(message.getTgtName(), message.getToState()); ``` might be cleaner. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
