[ 
https://issues.apache.org/jira/browse/HELIX-674?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16397917#comment-16397917
 ] 

ASF GitHub Bot commented on HELIX-674:
--------------------------------------

Github user jiajunwang commented on a diff in the pull request:

    https://github.com/apache/helix/pull/145#discussion_r174326641
  
    --- Diff: 
helix-core/src/main/java/org/apache/helix/util/WeightAwareRebalanceUtil.java ---
    @@ -0,0 +1,141 @@
    +package org.apache.helix.util;
    +
    +import org.apache.helix.HelixException;
    +import org.apache.helix.ZNRecord;
    +import org.apache.helix.api.config.RebalanceConfig;
    +import 
org.apache.helix.api.rebalancer.constraint.AbstractRebalanceHardConstraint;
    +import 
org.apache.helix.api.rebalancer.constraint.AbstractRebalanceSoftConstraint;
    +import org.apache.helix.controller.common.PartitionStateMap;
    +import org.apache.helix.controller.common.ResourcesStateMap;
    +import 
org.apache.helix.controller.rebalancer.strategy.ConstraintRebalanceStrategy;
    +import org.apache.helix.controller.stages.ClusterDataCache;
    +import org.apache.helix.model.*;
    +
    +import java.util.ArrayList;
    +import java.util.HashMap;
    +import java.util.List;
    +import java.util.Map;
    +
    +/**
    + * A rebalance tool that generate an resource partition assignment based 
on the input.
    + * Note the assignment won't be automatically applied to the cluster. 
Applications are supposed to
    + * apply the change.
    + */
    +public class WeightAwareRebalanceUtil {
    +  private final ClusterConfig _clusterConfig;
    +  private final Map<String, InstanceConfig> _instanceConfigMap = new 
HashMap<>();
    +  // For the possible customized state models.
    +  private final Map<String, StateModelDefinition> _stateModelDefs = new 
HashMap<>();
    +  private final ClusterDataCache _dataCache;
    +
    +  public enum RebalanceOption {
    +    INIT,
    +    REASSIGN
    +  }
    +
    +  public WeightAwareRebalanceUtil(ClusterConfig clusterConfig,
    +      List<InstanceConfig> instanceConfigs) {
    +    for (InstanceConfig instanceConfig : instanceConfigs) {
    +      _instanceConfigMap.put(instanceConfig.getInstanceName(), 
instanceConfig);
    +    }
    +    _clusterConfig = clusterConfig;
    +
    +    _dataCache = new ClusterDataCache();
    +    _dataCache.setInstanceConfigMap(_instanceConfigMap);
    +    _dataCache.setClusterConfig(_clusterConfig);
    +    List<LiveInstance> liveInstanceList = new ArrayList<>();
    +    for (String instance : _instanceConfigMap.keySet()) {
    +      LiveInstance liveInstance = new LiveInstance(instance);
    +      liveInstanceList.add(liveInstance);
    +    }
    +    _dataCache.setLiveInstances(liveInstanceList);
    +  }
    +
    +  /**
    +   * The method to generate partition assignment mappings.
    +   *
    +   * @param resourceConfigs    Config of all the resources that need to be 
rebalanced.
    +   *                           The tool throws Exception if any resource 
has no IS or broken/uninitialized IS.
    +   *                           The tool throws Exception if any resource 
is in full-auto mode.
    +   * @param existingAssignment The existing partition assignment of the 
input resources.
    +   * @param option             INIT or REASSIGN
    +   *                           INIT: Keep existing assignment. Only 
generate new partition assignment.
    --- End diff --
    
    And about the last point. You are correct. Since the rebalance algorithm is 
optimized to minimize movement, even with full rebalance, the new mapping is 
supposed to keep as much same assignment as the old one. Just not guaranteed.


> Constraint Based Resource Rebalancer
> ------------------------------------
>
>                 Key: HELIX-674
>                 URL: https://issues.apache.org/jira/browse/HELIX-674
>             Project: Apache Helix
>          Issue Type: New Feature
>            Reporter: Jiajun Wang
>            Assignee: Jiajun Wang
>            Priority: Major
>             Fix For: 0.8.x
>
>         Attachments: Constraint-BasedResourceRebalancing-080318-2226-240.pdf
>
>
> Helix rebalancer assigns resources according to different strategies. 
> Recently, we optimize the strategy for evenness and minimize movement. 
> However, the evenness here only applies to partition numbers. Moreover, we've 
> got more requests for customizable rebalancer from our users.
> Take partition weight as an example:
> In reality, partition replicas have different size. We use "partition weight" 
> as an abstraction of the partition size. It can be network traffic usage, 
> disk usage, or any other combined factors.
> Given each partition may have different weights, Helix should be able to 
> assign partition accordingly. So that the distribution would be even 
> regarding the weight.
> In this project, we are planning new rebalancer mechanism that generates 
> resource partition assignment according to a list of "constraints". Current 
> rebalance strategy can be regarded as one kind of constraint. Moving forward, 
> Helix users would be able to extend the constraint interface using their own 
> logic.
> Some init discussions are in progress and we will have a proposal posted here 
> soon.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to