HBASE-17707 New More Accurate Table Skew cost function/generator

This reverts commit 3b914df949.

Signed-off-by: tedyu <yuzhihong@gmail.com>

hbase-server/src/main/java/org/apache/hadoop/hbase/master/balancer/BaseLoadBalancer.java

@@ -53,6 +53,7 @@ import org.apache.hadoop.hbase.master.RackManager;
 import org.apache.hadoop.hbase.master.RegionPlan;
 import org.apache.hadoop.hbase.master.balancer.BaseLoadBalancer.Cluster.Action.Type;
 import org.apache.hadoop.hbase.security.access.AccessControlLists;
+import org.apache.hadoop.hbase.util.Pair;
 import org.apache.hadoop.util.StringUtils;
 
 import com.google.common.annotations.VisibleForTesting;
@@ -140,6 +141,7 @@ public abstract class BaseLoadBalancer implements LoadBalancer {
     int[]   initialRegionIndexToServerIndex;    //regionIndex -> serverIndex (initial cluster state)
     int[]   regionIndexToTableIndex;     //regionIndex -> tableIndex
     int[][] numRegionsPerServerPerTable; //serverIndex -> tableIndex -> # regions
+    int[]   numRegionsPerTable;          // tableIndex -> number of regions that table has
     int[]   numMaxRegionsPerTable;       //tableIndex -> max number of regions in a single RS
     int[]   regionIndexToPrimaryIndex;   //regionIndex -> regionIndex of the primary
     boolean hasRegionReplicas = false;   //whether there is regions with replicas
@@ -330,6 +332,7 @@ public abstract class BaseLoadBalancer implements LoadBalancer {
 
       numTables = tables.size();
       numRegionsPerServerPerTable = new int[numServers][numTables];
+      numRegionsPerTable = new int[numTables];
 
       for (int i = 0; i < numServers; i++) {
         for (int j = 0; j < numTables; j++) {
@@ -339,6 +342,7 @@ public abstract class BaseLoadBalancer implements LoadBalancer {
 
       for (int i=0; i < regionIndexToServerIndex.length; i++) {
         if (regionIndexToServerIndex[i] >= 0) {
+          numRegionsPerTable[regionIndexToTableIndex[i]]++;
           numRegionsPerServerPerTable[regionIndexToServerIndex[i]][regionIndexToTableIndex[i]]++;
         }
       }
@@ -470,6 +474,76 @@ public abstract class BaseLoadBalancer implements LoadBalancer {
       }
     }
 
+    /**
+     * Returns the minimum number of regions of a table T each server would store if T were
+     * perfectly distributed (i.e. round-robin-ed) across the cluster
+     */
+    public int minRegionsIfEvenlyDistributed(int table) {
+      return numRegionsPerTable[table] / numServers;
+    }
+
+    /**
+     * Returns the maximum number of regions of a table T each server would store if T were
+     * perfectly distributed (i.e. round-robin-ed) across the cluster
+     */
+    public int maxRegionsIfEvenlyDistributed(int table) {
+      int min = minRegionsIfEvenlyDistributed(table);
+      return numRegionsPerTable[table] % numServers == 0 ? min : min + 1;
+    }
+
+    /**
+     * Returns the number of servers that should hold maxRegionsIfEvenlyDistributed for a given
+     * table. A special case here is if maxRegionsIfEvenlyDistributed == minRegionsIfEvenlyDistributed,
+     * in which case all servers should hold the max
+     */
+    public int numServersWithMaxRegionsIfEvenlyDistributed(int table) {
+      int numWithMax = numRegionsPerTable[table] % numServers;
+      if (numWithMax == 0) {
+        return numServers;
+      } else {
+        return numWithMax;
+      }
+    }
+
+    /**
+     * Returns true iff at least one server in the cluster stores either more than the min/max load
+     * per server when all regions are evenly distributed across the cluster
+     */
+    public boolean hasUnevenRegionDistribution() {
+      int minLoad = numRegions / numServers;
+      int maxLoad = numRegions % numServers == 0 ? minLoad : minLoad + 1;
+      for (int server = 0; server < numServers; server++) {
+        int numRegions = getNumRegions(server);
+        if (numRegions > maxLoad || numRegions < minLoad) {
+          return true;
+        }
+      }
+      return false;
+    }
+
+    /**
+     * Returns a pair where the first server is that with the least number of regions across the
+     * cluster and the second server is that with the most number of regions across the cluster
+     */
+    public Pair<Integer, Integer> findLeastAndMostLoadedServers() {
+      int minServer = 0;
+      int maxServer = 0;
+      int minLoad = getNumRegions(minServer);
+      int maxLoad = minLoad;
+      for (int server = 1; server < numServers; server++) {
+        int numRegions = getNumRegions(server);
+        if (numRegions < minLoad) {
+          minServer = server;
+          minLoad = numRegions;
+        }
+        if (numRegions > maxLoad) {
+          maxServer = server;
+          maxLoad = numRegions;
+        }
+      }
+      return Pair.newPair(minServer, maxServer);
+    }
+
     /** An action to move or swap a region */
     public static class Action {
       public static enum Type {

hbase-server/src/main/java/org/apache/hadoop/hbase/master/balancer/StochasticLoadBalancer.java

@@ -18,10 +18,14 @@
 package org.apache.hadoop.hbase.master.balancer;
 
 import java.util.ArrayDeque;
+import java.util.ArrayList;
 import java.util.Arrays;
 import java.util.Collection;
+import java.util.Collections;
+import java.util.Comparator;
 import java.util.Deque;
 import java.util.HashMap;
+import java.util.Iterator;
 import java.util.LinkedList;
 import java.util.List;
 import java.util.Map;
@@ -30,7 +34,6 @@ import java.util.Random;
 
 import org.apache.commons.logging.Log;
 import org.apache.commons.logging.LogFactory;
-import org.apache.hadoop.hbase.classification.InterfaceAudience;
 import org.apache.hadoop.conf.Configuration;
 import org.apache.hadoop.hbase.ClusterStatus;
 import org.apache.hadoop.hbase.HBaseInterfaceAudience;
@@ -40,6 +43,7 @@ import org.apache.hadoop.hbase.RegionLoad;
 import org.apache.hadoop.hbase.ServerLoad;
 import org.apache.hadoop.hbase.ServerName;
 import org.apache.hadoop.hbase.TableName;
+import org.apache.hadoop.hbase.classification.InterfaceAudience;
 import org.apache.hadoop.hbase.master.MasterServices;
 import org.apache.hadoop.hbase.master.RegionPlan;
 import org.apache.hadoop.hbase.master.balancer.BaseLoadBalancer.Cluster.Action;
@@ -49,6 +53,10 @@ import org.apache.hadoop.hbase.master.balancer.BaseLoadBalancer.Cluster.MoveRegi
 import org.apache.hadoop.hbase.master.balancer.BaseLoadBalancer.Cluster.SwapRegionsAction;
 import org.apache.hadoop.hbase.util.Bytes;
 import org.apache.hadoop.hbase.util.EnvironmentEdgeManager;
+import org.apache.hadoop.hbase.util.Pair;
+
+import com.google.common.base.Optional;
+import com.google.common.base.Preconditions;
 
 /**
  * <p>This is a best effort load balancer. Given a Cost function F(C) =&gt; x It will
@@ -919,6 +927,225 @@ public class StochasticLoadBalancer extends BaseLoadBalancer {
     }
   }
 
+  /**
+   * Generates candidate actions to minimize the TableSkew cost function.
+   *
+   * For efficiency reasons, the cluster must be passed in when this generator is
+   * constructed. Every move generated is applied to the cost function
+   * (i.e. it is assumed that every action we generate is applied to the cluster).
+   * This means we can adjust our cost incrementally for the cluster, rather than
+   * recomputing at each iteration.
+   */
+  static class TableSkewCandidateGenerator extends CandidateGenerator {
+
+    // Mapping of table -> true iff too many servers in the cluster store at least
+    // cluster.maxRegionsIfEvenlydistributed(table)
+    boolean[] tablesWithEnoughServersWithMaxRegions = null;
+
+    @Override
+    Action generate(Cluster cluster) {
+      if (tablesWithEnoughServersWithMaxRegions == null || tablesWithEnoughServersWithMaxRegions.length != cluster.numTables) {
+        tablesWithEnoughServersWithMaxRegions = new boolean[cluster.numTables];
+      }
+      if (cluster.hasUnevenRegionDistribution()) {
+        Pair<Integer, Integer> leastAndMostLoadedServers = cluster.findLeastAndMostLoadedServers();
+        return moveFromTableWithEnoughRegions(cluster, leastAndMostLoadedServers.getSecond(), leastAndMostLoadedServers.getFirst());
+      } else {
+        Optional<TableAndServer> tableServer = findSkewedTableServer(cluster);
+        if (!tableServer.isPresent()) {
+          return Cluster.NullAction;
+        }
+        return findBestActionForTableServer(cluster, tableServer.get());
+      }
+    }
+
+    /**
+     * Returns a move fromServer -> toServer such that after the move fromServer will still have at least
+     * the min # regions in terms of table skew calculation
+     */
+    private Action moveFromTableWithEnoughRegions(Cluster cluster, int fromServer, int toServer) {
+      for (int table : getShuffledRangeOfInts(0, cluster.numTables)) {
+        int min = cluster.minRegionsIfEvenlyDistributed(table);
+        if (cluster.numRegionsPerServerPerTable[fromServer][table] > min) {
+          return getAction(fromServer, pickRandomRegionFromTableOnServer(cluster, fromServer, table), toServer, -1);
+        }
+      }
+      return Cluster.NullAction;
+    }
+
+    /**
+     * Picks a random subset of tables, then for each table T checks across cluster and returns first
+     * server (if any) which holds too many regions from T. Returns Optional.absent() if no servers
+     * are found that hold too many regions.
+     */
+    private Optional<TableAndServer> findSkewedTableServer(Cluster cluster) {
+      Optional<TableAndServer> tableServer = Optional.absent();
+      List<Integer> servers = getShuffledRangeOfInts(0, cluster.numServers);
+      Iterator<Integer> tableIter = getShuffledRangeOfInts(0, cluster.numTables).iterator();
+      while (tableIter.hasNext() && !tableServer.isPresent()) {
+        int table = tableIter.next();
+        int maxRegions = cluster.maxRegionsIfEvenlyDistributed(table);
+        int numShouldHaveMaxRegions = cluster.numServersWithMaxRegionsIfEvenlyDistributed(table);
+        int numWithMaxRegions = 0;
+        for (int server : servers) {
+          int numRegions = cluster.numRegionsPerServerPerTable[server][table];
+          // if more than max, server clearly has too many regions
+          if (numRegions > maxRegions) {
+            tableServer = Optional.of(new TableAndServer(table, server));
+            break;
+          }
+          // if equal to max, check to see if we are within acceptable limit
+          if (numRegions == maxRegions) {
+            numWithMaxRegions++;
+          }
+        }
+
+        tablesWithEnoughServersWithMaxRegions[table] = numWithMaxRegions >= numShouldHaveMaxRegions;
+        // If we have found a table with more than max, we are done
+        if (tableServer.isPresent()) {
+          break;
+        }
+
+        // Otherwise, check to see if there are too many servers with maxRegions
+        if (numWithMaxRegions > numShouldHaveMaxRegions) {
+          for (int server : servers) {
+            int numRegions = cluster.numRegionsPerServerPerTable[server][table];
+            if (numRegions == maxRegions) {
+              tableServer = Optional.of(new TableAndServer(table, server));
+              break;
+            }
+          }
+        }
+      }
+
+      return tableServer;
+    }
+
+    /**
+     * Returns an list of integers that stores [upper - lower] unique integers in random order
+     * s.t. for each integer i lower <= i < upper
+     */
+    private List<Integer> getShuffledRangeOfInts(int lower, int upper) {
+      Preconditions.checkArgument(lower < upper);
+      ArrayList<Integer> arr = new ArrayList<Integer>(upper - lower);
+      for (int i = lower; i < upper; i++) {
+        arr.add(i);
+      }
+      Collections.shuffle(arr);
+      return arr;
+    }
+
+    /**
+     * Pick a random region from the specified server and table. Returns -1 if no regions from
+     * the given table lie on the given server
+     */
+    protected int pickRandomRegionFromTableOnServer(Cluster cluster, int server, int table) {
+      if (server < 0 || table < 0) {
+        return -1;
+      }
+      List<Integer> regionsFromTable = new ArrayList<>();
+      for (int region : cluster.regionsPerServer[server]) {
+        if (cluster.regionIndexToTableIndex[region] == table) {
+          regionsFromTable.add(region);
+        }
+      }
+      return regionsFromTable.get(RANDOM.nextInt(regionsFromTable.size()));
+    }
+
+    /**
+     * Returns servers in the cluster that store fewer than k regions for the given table (sorted by
+     * servers with the fewest regions from givenTable first)
+     */
+    public List<Integer> getServersWithFewerThanKRegionsFromTable(final Cluster cluster, final int givenTable, int k) {
+      List<Integer> serversWithFewerThanK = new ArrayList<>();
+      for (int server = 0; server < cluster.numServers; server++) {
+        if (cluster.numRegionsPerServerPerTable[server][givenTable] < k) {
+          serversWithFewerThanK.add(server);
+        }
+      }
+      Collections.sort(serversWithFewerThanK, new Comparator<Integer>() {
+        @Override
+        public int compare(Integer o1, Integer o2) {
+          return cluster.numRegionsPerServerPerTable[o1.intValue()][givenTable] - cluster.numRegionsPerServerPerTable[o2.intValue()][givenTable];
+        }
+      });
+      return serversWithFewerThanK;
+    }
+
+    /**
+     * Given a table T for which server S stores too many regions, attempts to find a
+     * SWAP operation that will better balance the cluster
+     */
+    public Action findBestActionForTableServer(Cluster cluster, TableAndServer tableServer) {
+      int fromTable = tableServer.getTable();
+      int fromServer = tableServer.getServer();
+
+      int minNumRegions = cluster.minRegionsIfEvenlyDistributed(fromTable);
+      int maxNumRegions = cluster.maxRegionsIfEvenlyDistributed(fromTable);
+      List<Integer> servers;
+      if (tablesWithEnoughServersWithMaxRegions[fromTable]) {
+        servers = getServersWithFewerThanKRegionsFromTable(cluster, fromTable, minNumRegions);
+      } else {
+        servers = getServersWithFewerThanKRegionsFromTable(cluster, fromTable, maxNumRegions);
+      }
+
+      if (servers.isEmpty()) {
+        return Cluster.NullAction;
+      }
+
+      Optional<Action> swap = trySwap(cluster, fromServer, fromTable, servers);
+      if (swap.isPresent()) {
+        return swap.get();
+      }
+
+      // If we cannot perform a swap, we should do nothing
+      return Cluster.NullAction;
+    }
+
+    /**
+     * Given server1, table1, we try to find server2 and table2 such that
+     * at least 3 of the following 4 criteria are met
+     *
+     * 1) server1 has too many regions of table1
+     * 2) server1 has too few regions of table2
+     * 3) server2 has too many regions of table2
+     * 4) server2 has too few regions of table1
+     *
+     * We consider N regions from table T
+     *    too few if: N < cluster.minRegionsIfEvenlyDistributed(T)
+     *    too many if: N > cluster.maxRegionsIfEvenlyDistributed(T)
+     *
+     * Because (1) and (4) are true apriori, we only need to check for (2) and (3).
+     *
+     * If 3 of the 4 criteria are met, we return a swap operation between
+     * randomly selected regions from table1 on server1 and from table2 on server2.
+     *
+     * Optional.absent() is returned if we could not find such a SWAP.
+     */
+    private Optional<Action> trySwap(Cluster cluster, int server1, int table1, List<Integer> candidateServers) {
+      // Because conditions (1) and (4) are true apriori, we only need to meet one of conditions (2) or (3)
+      List<Integer> tables = getShuffledRangeOfInts(0, cluster.numTables);
+      for (int table2 : tables) {
+        int minRegions = cluster.minRegionsIfEvenlyDistributed(table2);
+        int maxRegions = cluster.maxRegionsIfEvenlyDistributed(table2);
+        for (int server2 : candidateServers) {
+          int numRegions1 = cluster.numRegionsPerServerPerTable[server1][table2];
+          int numRegions2 = cluster.numRegionsPerServerPerTable[server2][table2];
+          if (numRegions2 == 0) {
+            continue;
+          }
+          if ((numRegions1 < minRegions || numRegions2 > maxRegions) ||
+              (minRegions != maxRegions && numRegions1 == minRegions && numRegions2 == maxRegions)) {
+            int region1 = pickRandomRegionFromTableOnServer(cluster, server1, table1);
+            int region2 = pickRandomRegionFromTableOnServer(cluster, server2, table2);
+            return Optional.of(getAction(server1, region1, server2, region2));
+          }
+        }
+      }
+      return Optional.absent();
+    }
+  }
+
   /**
    * Base class of StochasticLoadBalancer's Cost Functions.
    */
@@ -966,8 +1193,7 @@ public class StochasticLoadBalancer extends BaseLoadBalancer {
         break;
       case SWAP_REGIONS:
         SwapRegionsAction a = (SwapRegionsAction) action;
-        regionMoved(a.fromRegion, a.fromServer, a.toServer);
+        regionSwapped(a.fromRegion, a.fromServer, a.toRegion, a.toServer);
-        regionMoved(a.toRegion, a.toServer, a.fromServer);
         break;
       default:
         throw new RuntimeException("Uknown action:" + action.type);
@@ -977,6 +1203,11 @@ public class StochasticLoadBalancer extends BaseLoadBalancer {
     protected void regionMoved(int region, int oldServer, int newServer) {
     }
 
+    protected void regionSwapped(int region1, int server1, int region2, int server2) {
+      regionMoved(region1, server1, server2);
+      regionMoved(region2, server2, server1);
+    }
+
     abstract double cost();
 
     /**
@@ -1170,9 +1401,188 @@ public class StochasticLoadBalancer extends BaseLoadBalancer {
         "hbase.master.balancer.stochastic.tableSkewCost";
     private static final float DEFAULT_TABLE_SKEW_COST = 35;
 
+    /**
+     * Ranges from 0.0 to 1.0 and is the proportion of how much the most skewed table
+     * (as opposed to the average skew across all tables) should affect TableSkew cost
+     */
+    private static final String MAX_TABLE_SKEW_WEIGHT_KEY =
+        "hbase.master.balancer.stochastic.maxTableSkewWeight";
+    private float DEFAULT_MAX_TABLE_SKEW_WEIGHT = 0.0f;
+
+    private final float maxTableSkewWeight;
+    private final float avgTableSkewWeight;
+
+    // Number of moves for each table required to bring the cluster to a perfectly balanced
+    // state (i.e. as if you had round-robin-ed regions across cluster)
+    private int[] numMovesPerTable;
+
     TableSkewCostFunction(Configuration conf) {
       super(conf);
       this.setMultiplier(conf.getFloat(TABLE_SKEW_COST_KEY, DEFAULT_TABLE_SKEW_COST));
+      maxTableSkewWeight = conf.getFloat(MAX_TABLE_SKEW_WEIGHT_KEY, DEFAULT_MAX_TABLE_SKEW_WEIGHT);
+      Preconditions.checkArgument(0.0 <= maxTableSkewWeight && maxTableSkewWeight <= 1.0);
+      avgTableSkewWeight = 1 - maxTableSkewWeight;
+    }
+
+    /**
+     * Computes cost by:
+     *
+     * 1) Computing a skew score for each table (based on the number of regions
+     * from that table that would have to be moved to reach an evenly balanced state)
+     *
+     * 2) Taking a weighted average of the highest skew score with the average skew score
+     *
+     * 3) Square rooting that value to more evenly distribute the values between 0-1
+     * (since we have observed they are generally very small).
+     *
+     * @return the table skew cost for the cluster
+     */
+    @Override
+    double cost() {
+      double[] skewPerTable = computeSkewPerTable();
+      if (skewPerTable.length == 0) {
+        return 0;
+      }
+      double maxTableSkew = max(skewPerTable);
+      double avgTableSkew = average(skewPerTable);
+
+      return Math.sqrt(maxTableSkewWeight * maxTableSkew + avgTableSkewWeight * avgTableSkew);
+    }
+
+    @Override
+    void init(Cluster cluster) {
+      super.init(cluster);
+      numMovesPerTable = computeNumMovesPerTable();
+    }
+
+    /**
+     * Adjusts computed number of moves after two regions have been swapped
+     */
+    @Override
+    protected void regionSwapped(int region1, int server1, int region2, int server2) {
+      // If different tables, simply perform two moves
+      if (cluster.regionIndexToTableIndex[region1] != cluster.regionIndexToTableIndex[region2]) {
+        super.regionSwapped(region1, server1, region2, server2);
+        return;
+      }
+      // If same table, do nothing
+    }
+
+    /**
+     * Adjusts computed number of moves per table after a region has been moved
+     */
+    @Override
+    protected void regionMoved(int region, int oldServer, int newServer) {
+      int table = cluster.regionIndexToTableIndex[region];
+      numMovesPerTable[table] = computeNumMovesForTable(table);
+    }
+
+    /**
+     * Returns a mapping of table -> numMoves, where numMoves is the number of regions required to bring
+     * each table to a fully balanced state (i.e. as if its regions had been round-robin-ed across the cluster).
+     */
+    private int[] computeNumMovesPerTable() {
+      // Determine # region moves required for each table to have regions perfectly distributed across cluster
+      int[] numMovesPerTable = new int[cluster.numTables];
+      for (int table = 0; table < cluster.numTables; table++) {
+        numMovesPerTable[table] = computeNumMovesForTable(table);
+      }
+      return numMovesPerTable;
+    }
+
+    /**
+     * Computes the number of moves required across all servers to bring the given table to a balanced state
+     * (i.e. as if its regions had been round-robin-ed across the cluster). We only consider moves as # of regions
+     * that need to be sent, not received, so that we do not double count region moves.
+     */
+    private int computeNumMovesForTable(int table) {
+      int numMinRegions = cluster.minRegionsIfEvenlyDistributed(table);
+      int numMaxRegions = cluster.maxRegionsIfEvenlyDistributed(table);
+      int numMaxServersRemaining = cluster.numServersWithMaxRegionsIfEvenlyDistributed(table);
+      int numMoves = 0;
+
+      for (int server = 0; server < cluster.numServers; server++) {
+        int numRegions = cluster.numRegionsPerServerPerTable[server][table];
+        if (numRegions >= numMaxRegions && numMaxServersRemaining > 0) {
+          numMoves += numRegions - numMaxRegions;
+          numMaxServersRemaining--;
+        } else if (numRegions > numMinRegions) {
+          numMoves += numRegions - numMinRegions;
+        }
+      }
+      return numMoves;
+    }
+
+    /**
+     * Returns mapping of tableIndex -> tableSkewScore, where tableSkewScore is a double between 0 to 1 with
+     * 0 indicating no table skew (i.e. perfect distribution of regions among servers), and 1 representing
+     * pathological table skew (i.e. all of a servers regions belonging to one table).
+     */
+    private double[] computeSkewPerTable() {
+      if (numMovesPerTable == null) {
+        numMovesPerTable = computeNumMovesPerTable();
+      }
+      double[] scaledSkewPerTable = new double[numMovesPerTable.length];
+      for (int table = 0; table < numMovesPerTable.length; table++) {
+        int numTotalRegions = cluster.numRegionsPerTable[table];
+        int maxRegions = cluster.maxRegionsIfEvenlyDistributed(table);
+        int pathologicalNumMoves = numTotalRegions - maxRegions;
+        scaledSkewPerTable[table] = pathologicalNumMoves == 0 ? 0 : (double) numMovesPerTable[table] / pathologicalNumMoves;
+      }
+      return scaledSkewPerTable;
+    }
+
+    /**
+     * Returns the max of the values in the passed array
+     */
+    private double max(double[] arr) {
+      double max = arr[0];
+      for (double d : arr) {
+        if (d > max) {
+          max = d;
+        }
+      }
+      return max;
+    }
+
+    /**
+     * Returns the average of the values in the passed array
+     */
+    private double average(double[] arr) {
+      double sum = 0;
+      for (double d : arr) {
+        sum += d;
+      }
+      return sum / arr.length;
+    }
+  }
+
+  /**
+   * Compute the cost of a potential cluster configuration based upon how evenly
+   * distributed tables are.
+   *
+   * @deprecated replaced by TableSkewCostFunction
+   * This function only considers the maximum # of regions of each table stored
+   * on any one server. This, however, neglects a number of cases. Consider the case
+   * where N servers store 1 more region than as if the regions had been round robin-ed
+   * across the cluster, but then K servers stored 0 regions of the table. The maximum
+   * # regions stored would not properly reflect the table-skew of the cluster.
+   *
+   * Furthermore, this relies upon the cluster.numMaxRegionsPerTable field, which is not
+   * properly updated. The values per table only increase as the cluster shifts (i.e.
+   * as new maxima are found), but they do not go down when the maximum skew decreases
+   * for a particular table.
+   */
+  @Deprecated
+  static class OldTableSkewCostFunction extends CostFunction {
+
+    private static final String TABLE_SKEW_COST_KEY =
+        "hbase.master.balancer.stochastic.tableSkewCost";
+    private static final float DEFAULT_TABLE_SKEW_COST = 35;
+
+    OldTableSkewCostFunction(Configuration conf) {
+      super(conf);
+      this.setMultiplier(conf.getFloat(TABLE_SKEW_COST_KEY, DEFAULT_TABLE_SKEW_COST));
     }
 
     @Override
@@ -1588,10 +1998,32 @@ public class StochasticLoadBalancer extends BaseLoadBalancer {
     }
   }
 
+  /**
+   * Data structure that holds table and server indexes
+   */
+  static class TableAndServer {
+    private final int table;
+    private final int server;
+
+    public TableAndServer(int table, int server) {
+      this.table = table;
+      this.server = server;
+    }
+
+    public int getTable() {
+      return table;
+    }
+
+    public int getServer() {
+      return server;
+    }
+  }
+
   /**
    * A helper function to compose the attribute name from tablename and costfunction name
    */
   public static String composeAttributeName(String tableName, String costFunctionName) {
     return tableName + TABLE_FUNCTION_SEP + costFunctionName;
   }
+
 }

hbase-server/src/test/java/org/apache/hadoop/hbase/master/balancer/TestStochasticLoadBalancer.java

@@ -48,6 +48,8 @@ import org.apache.hadoop.hbase.master.MockNoopMasterServices;
 import org.apache.hadoop.hbase.master.RackManager;
 import org.apache.hadoop.hbase.master.RegionPlan;
 import org.apache.hadoop.hbase.master.balancer.BaseLoadBalancer.Cluster;
+import org.apache.hadoop.hbase.master.balancer.StochasticLoadBalancer.CandidateGenerator;
+import org.apache.hadoop.hbase.master.balancer.StochasticLoadBalancer.TableSkewCandidateGenerator;
 import org.apache.hadoop.hbase.testclassification.FlakeyTests;
 import org.apache.hadoop.hbase.testclassification.MediumTests;
 import org.apache.hadoop.hbase.util.Bytes;
@@ -119,7 +121,9 @@ public class TestStochasticLoadBalancer extends BalancerTestBase {
    */
   @Test
   public void testBalanceCluster() throws Exception {
-
+    float oldMinCostNeedBalance = conf.getFloat(StochasticLoadBalancer.MIN_COST_NEED_BALANCE_KEY, 0.05f);
+    conf.setFloat(StochasticLoadBalancer.MIN_COST_NEED_BALANCE_KEY, 0.02f);
+    loadBalancer.setConf(conf);
     for (int[] mockCluster : clusterStateMocks) {
       Map<ServerName, List<HRegionInfo>> servers = mockClusterServers(mockCluster);
       List<ServerAndLoad> list = convertToList(servers);
@@ -135,6 +139,9 @@ public class TestStochasticLoadBalancer extends BalancerTestBase {
         returnServer(entry.getKey());
       }
     }
+    // reset config
+    conf.setFloat(StochasticLoadBalancer.MIN_COST_NEED_BALANCE_KEY, oldMinCostNeedBalance);
+    loadBalancer.setConf(conf);
   }
 
   @Test
@@ -255,6 +262,32 @@ public class TestStochasticLoadBalancer extends BalancerTestBase {
     assertEquals(2.5, result, 0.01);
  }
 
+  @Test (timeout=60000)
+  public void testTableSkewCandidateGeneratorConvergesToZero() {
+    int replication = 1;
+    StochasticLoadBalancer.CostFunction
+        costFunction = new StochasticLoadBalancer.TableSkewCostFunction(conf);
+    CandidateGenerator generator = new TableSkewCandidateGenerator();
+    for (int i = 0; i < 100; i++) {
+      int numNodes = rand.nextInt(500) + 1; // num nodes between 1 - 500
+      int numTables = rand.nextInt(500) + 1; // num tables between 1 and 1000
+      int numRegions = rand.nextInt(numTables * 99) + Math.max(numTables, numNodes); // num regions between max(numTables, numNodes) - numTables*100
+      int numRegionsPerServer = rand.nextInt(numRegions / numNodes) + 1; // num regions per server (except one) between 1 and numRegions / numNodes
+
+      Map<ServerName, List<HRegionInfo>> serverMap = createServerMap(numNodes, numRegions, numRegionsPerServer, replication, numTables);
+      BaseLoadBalancer.Cluster cluster = new Cluster(serverMap, null, null, null);
+      costFunction.init(cluster);
+      double cost = costFunction.cost();
+      while (cost > 0) {
+        Cluster.Action action = generator.generate(cluster);
+        cluster.doAction(action);
+        costFunction.postAction(action);
+        cost = costFunction.cost();
+      }
+      assertEquals(0, cost, .000000000001);
+    }
+  }
+
   @Test
   public void testCostFromArray() {
     Configuration conf = HBaseConfiguration.create();

hbase-server/src/test/java/org/apache/hadoop/hbase/master/balancer/TestStochasticLoadBalancer2.java

@@ -35,6 +35,7 @@ public class TestStochasticLoadBalancer2 extends BalancerTestBase {
     conf.setFloat("hbase.master.balancer.stochastic.maxMovePercent", 1.0f);
     conf.setLong(StochasticLoadBalancer.MAX_STEPS_KEY, 2000000L);
     conf.setFloat("hbase.master.balancer.stochastic.localityCost", 0);
+
     conf.setLong("hbase.master.balancer.stochastic.maxRunningTime", 90 * 1000); // 90 sec
     conf.setFloat("hbase.master.balancer.stochastic.minCostNeedBalance", 0.05f);
     loadBalancer.setConf(conf);
@@ -70,6 +71,7 @@ public class TestStochasticLoadBalancer2 extends BalancerTestBase {
   public void testRegionReplicasOnMidClusterHighReplication() {
     conf.setLong(StochasticLoadBalancer.MAX_STEPS_KEY, 4000000L);
     conf.setLong("hbase.master.balancer.stochastic.maxRunningTime", 120 * 1000); // 120 sec
+    conf.setFloat("hbase.master.balancer.stochastic.tableSkewCost", 4);
     loadBalancer.setConf(conf);
     int numNodes = 80;
     int numRegions = 6 * numNodes;
@@ -77,6 +79,8 @@ public class TestStochasticLoadBalancer2 extends BalancerTestBase {
     int numRegionsPerServer = 5;
     int numTables = 10;
     testWithCluster(numNodes, numRegions, numRegionsPerServer, replication, numTables, false, true);
+    // reset config
+    conf.setFloat("hbase.master.balancer.stochastic.tableSkewCost", 35);
   }
 
   @Test (timeout = 800000)