|
| 1 | +/* |
| 2 | + * Licensed to the Apache Software Foundation (ASF) under one |
| 3 | + * or more contributor license agreements. See the NOTICE file |
| 4 | + * distributed with this work for additional information |
| 5 | + * regarding copyright ownership. The ASF licenses this file |
| 6 | + * to you under the Apache License, Version 2.0 (the |
| 7 | + * "License"); you may not use this file except in compliance |
| 8 | + * with the License. You may obtain a copy of the License at |
| 9 | + * |
| 10 | + * http://www.apache.org/licenses/LICENSE-2.0 |
| 11 | + * |
| 12 | + * Unless required by applicable law or agreed to in writing, software |
| 13 | + * distributed under the License is distributed on an "AS IS" BASIS, |
| 14 | + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 15 | + * See the License for the specific language governing permissions and |
| 16 | + * limitations under the License. |
| 17 | + */ |
| 18 | +package com.datastax.oss.driver.internal.core.loadbalancing; |
| 19 | + |
| 20 | +import static java.util.concurrent.TimeUnit.SECONDS; |
| 21 | + |
| 22 | +import com.datastax.oss.driver.api.core.config.DefaultDriverOption; |
| 23 | +import com.datastax.oss.driver.api.core.context.DriverContext; |
| 24 | +import com.datastax.oss.driver.api.core.metadata.Node; |
| 25 | +import com.datastax.oss.driver.api.core.session.Request; |
| 26 | +import com.datastax.oss.driver.api.core.session.Session; |
| 27 | +import com.datastax.oss.driver.internal.core.util.ArrayUtils; |
| 28 | +import com.datastax.oss.driver.internal.core.util.collection.QueryPlan; |
| 29 | +import com.datastax.oss.driver.internal.core.util.collection.SimpleQueryPlan; |
| 30 | +import edu.umd.cs.findbugs.annotations.NonNull; |
| 31 | +import java.util.ArrayList; |
| 32 | +import java.util.Comparator; |
| 33 | +import java.util.List; |
| 34 | +import java.util.Objects; |
| 35 | +import java.util.Queue; |
| 36 | +import java.util.Random; |
| 37 | +import java.util.Set; |
| 38 | +import java.util.concurrent.ThreadLocalRandom; |
| 39 | +import org.slf4j.Logger; |
| 40 | +import org.slf4j.LoggerFactory; |
| 41 | + |
| 42 | +/** |
| 43 | + * A load balancing policy that optimally balances between sending load to local token holder, |
| 44 | + * rack replicas, and local datacenter replicas (in that order). |
| 45 | + * |
| 46 | + * The default weights are good for the vast majority of use cases, but you can tweak them to get different behavior. |
| 47 | + */ |
| 48 | +public class WeightedLoadBalancingPolicy extends DefaultLoadBalancingPolicy { |
| 49 | + private static final Logger LOG = |
| 50 | + LoggerFactory.getLogger(WeightedLoadBalancingPolicy.class); |
| 51 | + // Each client will randomly skew so traffic is introduced gradually to a newly up replica |
| 52 | + // Each client will start sending at a period between 15s and 30, and they will gradually |
| 53 | + // increase load over the next 15 seconds. |
| 54 | + private static final long DELAY_TRAFFIC_SKEW_MILLIS = SECONDS.toMillis(15); |
| 55 | + private static final long DELAY_TRAFFIC_MILLIS = |
| 56 | + DELAY_TRAFFIC_SKEW_MILLIS + ThreadLocalRandom.current().nextLong(DELAY_TRAFFIC_SKEW_MILLIS); |
| 57 | + |
| 58 | + // By default we will only score this many nodes, the rest will get added on without scoring. |
| 59 | + // We don't usually need to score every single node if there are more than a few. |
| 60 | + static final int DEFAULT_SCORED_PLAN_SIZE = 8; |
| 61 | + // Default multiplicative weights. Think of this like "How much concurrency must occur |
| 62 | + // before I fail off this node to the next". Note that these defaults are intentionally |
| 63 | + // meant to shed load to unloaded rack coordinators when a replica set is all |
| 64 | + // relatively heavily loaded (specifically 3x as loaded). |
| 65 | + static final double DEFAULT_WEIGHT_NON_RACK = 4.0; |
| 66 | + static final double DEFAULT_WEIGHT_NON_REPLICA = 12.0; |
| 67 | + static final double DEFAULT_WEIGHT_STARTING = 16.0; |
| 68 | + static final double DEFAULT_WEIGHT_UNHEALTHY = 64.0; |
| 69 | + |
| 70 | + private final int planSize; |
| 71 | + private final double weightNonRack; |
| 72 | + private final double weightNonReplica; |
| 73 | + private final double weightStarting; |
| 74 | + private final double weightUnhealthy; |
| 75 | + |
| 76 | + public WeightedLoadBalancingPolicy( |
| 77 | + @NonNull DriverContext context, |
| 78 | + @NonNull String profileName) { |
| 79 | + super(context, profileName); |
| 80 | + this.planSize = profile.getInt(DefaultDriverOption.LOAD_BALANCING_SCORED_PLAN_SIZE, DEFAULT_SCORED_PLAN_SIZE); |
| 81 | + // Choices of weights will change how this load balancer prefers endpoints. |
| 82 | + // The weight is relative to the outstanding concurrency. |
| 83 | + this.weightNonRack = profile.getDouble( |
| 84 | + DefaultDriverOption.LOAD_BALANCING_WEIGHT_NON_RACK, DEFAULT_WEIGHT_NON_RACK); |
| 85 | + this.weightNonReplica = profile.getDouble( |
| 86 | + DefaultDriverOption.LOAD_BALANCING_WEIGHT_NON_REPLICA, DEFAULT_WEIGHT_NON_REPLICA); |
| 87 | + this.weightStarting = profile.getDouble( |
| 88 | + DefaultDriverOption.LOAD_BALANCING_WEIGHT_STARTING, DEFAULT_WEIGHT_STARTING); |
| 89 | + this.weightUnhealthy = profile.getDouble( |
| 90 | + DefaultDriverOption.LOAD_BALANCING_WEIGHT_UNHEALTHY, DEFAULT_WEIGHT_UNHEALTHY); |
| 91 | + } |
| 92 | + |
| 93 | + @NonNull |
| 94 | + @Override |
| 95 | + public Queue<Node> newQueryPlan(Request request, Session session) { |
| 96 | + if (session == null) { |
| 97 | + return super.newQueryPlan(request, null); |
| 98 | + } |
| 99 | + |
| 100 | + // Take a copy of nodes and reference to replicas since the node map is concurrent |
| 101 | + Set<Node> dcNodeSet = getLiveNodes().dc(getLocalDatacenter()); |
| 102 | + Set<Node> replicaSet = getReplicas(request, session); |
| 103 | + |
| 104 | + long nowNanos = nanoTime(); |
| 105 | + long nowMillis = milliTime(); |
| 106 | + |
| 107 | + // collect known replica nodes |
| 108 | + List<NodeScore> nodeScores = new ArrayList<>(this.planSize); |
| 109 | + for (Node replicaNode : replicaSet) { |
| 110 | + if (dcNodeSet.contains(replicaNode)) { |
| 111 | + nodeScores.add(new NodeScore(replicaNode, |
| 112 | + getWeightedScore(replicaNode, session, nowMillis, nowNanos, true))); |
| 113 | + |
| 114 | + if (nodeScores.size() == this.planSize) { |
| 115 | + break; // TODO (akhaku) add the rest of the nodes once we port the tests to OSS |
| 116 | + } |
| 117 | + } |
| 118 | + } |
| 119 | + |
| 120 | + // collect any non-replicas, if we need to and there are some available |
| 121 | + if (nodeScores.size() < this.planSize && nodeScores.size() < dcNodeSet.size()) { |
| 122 | + Random rand = getRandom(); |
| 123 | + final Node[] dcNodes = dcNodeSet.toArray(new Node[0]); |
| 124 | + |
| 125 | + for (int i = 0; i < dcNodes.length; i++) { |
| 126 | + // pick a random target; shuffle it up front, so we don't revisit |
| 127 | + int nextIndex = i + rand.nextInt(dcNodes.length - i); |
| 128 | + ArrayUtils.swap(dcNodes, i, nextIndex); |
| 129 | + Node dcNode = dcNodes[i]; |
| 130 | + |
| 131 | + // skip replicas; they were already inserted |
| 132 | + // otherwise, found a valid node: score it |
| 133 | + if (!replicaSet.contains(dcNode)) { |
| 134 | + nodeScores.add(new NodeScore(dcNode, |
| 135 | + getWeightedScore(dcNode, session, nowMillis, nowNanos, false))); |
| 136 | + |
| 137 | + // if we scored, we might by now have already scored enough of what we need |
| 138 | + if (nodeScores.size() == this.planSize || nodeScores.size() == dcNodes.length) { |
| 139 | + break; |
| 140 | + } |
| 141 | + } |
| 142 | + } |
| 143 | + |
| 144 | + // by now we've scored everything we need to meet planSize, or if not, at least everything available |
| 145 | + } |
| 146 | + |
| 147 | + // At this point we have a small, typically 8 element array containing all local |
| 148 | + // datacenter replicas and potentially some random choices from the rest of the datacenter. |
| 149 | + // |
| 150 | + // We now rank nodes by a score function that takes into account outstanding requests weighted |
| 151 | + // by replica status, rack placement, uptime, and health status. In general, we expect to |
| 152 | + // see the following order |
| 153 | + // 1. Rack replicas |
| 154 | + // 2. Datacenter replicas |
| 155 | + // 3. Rack nodes |
| 156 | + // 4. Datacenter nodes |
| 157 | + // We expect these orderings to move around when nodes are overloaded. For example if the |
| 158 | + // local zone replica has too much load we will failover to other replicas. If those |
| 159 | + // replicas are too slow we will failover to other rack nodes. |
| 160 | + |
| 161 | + // sort, extract, convert to query plan |
| 162 | + nodeScores.sort(Comparator.comparingDouble(NodeScore::getScore)); |
| 163 | + Node[] candidate = new Node[nodeScores.size()]; |
| 164 | + for (int i = 0; i < candidate.length; i++) { |
| 165 | + candidate[i] = nodeScores.get(i).getNode(); |
| 166 | + } |
| 167 | + |
| 168 | + QueryPlan queryPlan = candidate.length == 0 ? QueryPlan.EMPTY : new SimpleQueryPlan((Object[]) candidate); |
| 169 | + return maybeAddDcFailover(request, queryPlan); |
| 170 | + } |
| 171 | + |
| 172 | + protected String getLocalRack() { |
| 173 | + return ""; // TODO (akhaku) internally we passed it through the context, for OSS perhaps something like the local DC helper? |
| 174 | + } |
| 175 | + |
| 176 | + protected boolean inRack(Node node) { |
| 177 | + if (node == null || node.getRack() == null) return false; |
| 178 | + return node.getRack().equals(this.getLocalRack()); |
| 179 | + } |
| 180 | + |
| 181 | + protected double getWeightedScore(Node node, Session session, long nowMillis, long nowNanos, boolean isReplica) { |
| 182 | + int base = Math.min(32768, 1 + getInFlight(node, session)); |
| 183 | + double weight = 1.0; |
| 184 | + |
| 185 | + if (!inRack(node)) weight *= this.weightNonRack; // 4.0 |
| 186 | + if (!isReplica) weight *= this.weightNonReplica; // 12.0 |
| 187 | + if (isUnhealthy(node, session, nowNanos)) weight *= this.weightUnhealthy; // 64.0 |
| 188 | + |
| 189 | + // We want to gradually ramp up traffic, shedding heavily at first and then allowing it back |
| 190 | + // in gradually. Note: |
| 191 | + // |
| 192 | + // 1. We cannot use nowNanos for this since node.getUpSinceMillis uses |
| 193 | + // System.currentTimeMillis (which may have no relation to nano time). |
| 194 | + // 2. getUpSinceMillis might return 0 or -1, in either case don't consider it freshly up. |
| 195 | + // 3. When a client starts up everything will appear to be freshly up, which is fine since |
| 196 | + // all nodes will randomly be shuffled to the front and back. |
| 197 | + long millisSinceUp = nowMillis - node.getUpSinceMillis(); |
| 198 | + if (millisSinceUp < (DELAY_TRAFFIC_MILLIS + DELAY_TRAFFIC_SKEW_MILLIS)) { |
| 199 | + double pShed = 1.0 - ((double) millisSinceUp / (DELAY_TRAFFIC_MILLIS + DELAY_TRAFFIC_SKEW_MILLIS)); |
| 200 | + if (pShed > getRandom().nextDouble()) { |
| 201 | + if (LOG.isTraceEnabled()) { |
| 202 | + String shed = String.format("%.2f", pShed); |
| 203 | + LOG.trace("[{}] shedding at startup [pShed={}, millisSinceUp={}]", node, shed, millisSinceUp); |
| 204 | + } |
| 205 | + weight *= this.weightStarting; // 16.0 |
| 206 | + } |
| 207 | + } |
| 208 | + |
| 209 | + double score = base * weight; |
| 210 | + if (LOG.isDebugEnabled()) { |
| 211 | + String msg = String.format("score=%.2f [base=%d, weight=%.2f]", score, base, weight); |
| 212 | + LOG.debug("[{}] {}", node, msg); |
| 213 | + } |
| 214 | + return score; |
| 215 | + } |
| 216 | + |
| 217 | + protected long milliTime() { |
| 218 | + return System.currentTimeMillis(); |
| 219 | + } |
| 220 | + |
| 221 | + protected Random getRandom() { |
| 222 | + return ThreadLocalRandom.current(); |
| 223 | + } |
| 224 | + |
| 225 | + /** |
| 226 | + * Wraps a Node alongside its score. |
| 227 | + * |
| 228 | + * Calculating scores is expensive, and not stable (could vary). By wrapping them we can be sure the score |
| 229 | + * is calculated only once and does not change during processing. |
| 230 | + */ |
| 231 | + static class NodeScore { |
| 232 | + final double score; |
| 233 | + final Node node; |
| 234 | + |
| 235 | + public NodeScore(Node node, double score) { |
| 236 | + this.node = node; |
| 237 | + this.score = score; |
| 238 | + } |
| 239 | + |
| 240 | + public Node getNode() { |
| 241 | + return node; |
| 242 | + } |
| 243 | + |
| 244 | + public double getScore() { |
| 245 | + return score; |
| 246 | + } |
| 247 | + |
| 248 | + @Override |
| 249 | + public boolean equals(Object o) { |
| 250 | + if (this == o) { |
| 251 | + return true; |
| 252 | + } |
| 253 | + if (o == null || getClass() != o.getClass()) { |
| 254 | + return false; |
| 255 | + } |
| 256 | + NodeScore nodeScore = (NodeScore) o; |
| 257 | + return Double.compare(score, nodeScore.score) == 0 && Objects.equals(node, nodeScore.node); |
| 258 | + } |
| 259 | + |
| 260 | + @Override |
| 261 | + public int hashCode() { |
| 262 | + return Objects.hash(score, node); |
| 263 | + } |
| 264 | + } |
| 265 | +} |
0 commit comments