Examples with TrackerClient com.linkedin.d2.balancer.clients.TrackerClient used on opensource projects

Example 86 with TrackerClient

use of com.linkedin.d2.balancer.clients.TrackerClient in project rest.li by linkedin.

the class DegraderLoadBalancerStrategyV2 method updateState.

/**
   * updateState
   *
   * We have two mechanisms to influence the health and traffic patterns of the client. They are
   * by load balancing (switching traffic from one host to another) and by degrading service
   * (dropping calls). We load balance by allocating points in a consistent hash ring based on the
   * computedDropRate of the individual TrackerClients, which takes into account the latency
   * seen by that TrackerClient's requests. We can alternatively, if the cluster is
   * unhealthy (by using a high latency watermark) drop a portion of traffic across all tracker
   * clients corresponding to this cluster.
   *
   * The reason we do not currently consider error rate when adjusting the hash ring is that
   * there are legitimate errors that servers can send back for clients to handle, such as
   * 400 return codes. A potential improvement would be to catch transport level exceptions and 500
   * level return codes, but the implication of that would need to be carefully understood and documented.
   *
   * We don't want both to reduce hash points and allow clients to manage their own drop rates
   * because the clients do not have a global view that the load balancing strategy does. Without
   * a global view, the clients won't know if it already has a reduced number of hash points. If the
   * client continues to drop at the same drop rate as before their points have been reduced, then
   * the client would have its outbound request reduced by both reduction in points and the client's
   * drop rate. To avoid this, the drop rate is managed globally by the load balancing strategy and
   * provided to each client. The strategy will ALTERNATE between adjusting the hash ring points or
   * the global drop rate in order to avoid double penalizing a client. See below:
   *
   * Period 1
   * We found the average latency is greater than high water mark.
   * Then increase the global drop rate for this cluster (let's say from 0% to 20%)
   * so 20% of all calls gets dropped.
   * .
   * .
   * Period 2
   * The average latency is still higher than high water mark and we found
   * it is especially high for few specific clients in the cluster
   * Then reduce the number of hash points for those clients in the hash ring, with the hope we'll
   * redirect the traffic to "healthier" client and reduce the average latency
   * .
   * .
   * Period 3
   * The average latency is still higher than high water mark
   * Then we will alternate strategy by increasing the global rate for the whole cluster again
   * .
   * .
   * repeat until the latency becomes smaller than high water mark and higher than low water mark
   * to maintain the state. If the latency becomes lower than low water mark that means the cluster
   * is getting healthier so we can serve more traffic so we'll start recovery as explained below
   *
   * We also have a mechanism for recovery if the number of points in the hash ring is not
   * enough to receive traffic. The initialRecoveryLevel is a number between 0.0 and 1.0, and
   * corresponds to a weight of the tracker client's full hash points. e.g. if a client
   * has a default 100 hash points in a ring, 0.0 means there's 0 point for the client in the ring
   * and 1.0 means there are 100 points in the ring for the client.
   * The second configuration, rampFactor, will geometrically increase the
   * previous recoveryLevel if traffic still hasn't been seen for that tracker client.
   *
   * The reason for using weight instead of real points is to allow an initialRecoveryLevel that corresponds to
   * less than one hash point. This would be useful if a "cooling off" period is desirable for the
   * misbehaving tracker clients i.e. given a full weight of 100 hash points, 0.005 initialRecoverylevel
   * 0 hashpoints at start and rampFactor = 2 means that there will be one cooling off period before the
   * client is reintroduced into the hash ring (see below).
   *
   * Period 1
   * 100 * 0.005 = 0.5 point -> So nothing in the hashring
   *
   * Period 2
   * 100 * (0.005 * 2 because of rampfactor) = 1 point -> So we'll add one point in the hashring
   *
   * Another example, given initialRecoveryLevel = 0.01, rampFactor = 2, and default tracker client hash
   * points of 100, we will increase the hash points in this pattern on successive update States:
   * 0.01, 0.02, 0.04, 0.08, 0.16, 0.32, etc. -> 1, 2, 4, 8, 16, 32 points in the hashring and aborting
   * as soon as calls are recorded for that tracker client.
   *
   * We also have highWaterMark and lowWaterMark as properties of the DegraderLoadBalancer strategy
   * so that the strategy can make decisions on whether to start dropping traffic GLOBALLY across
   * all tracker clients for this cluster. The amount of traffic to drop is controlled by the
   * globalStepUp and globalStepDown properties, where globalStepUp controls how much the global
   * drop rate increases per interval, and globalStepDown controls how much the global drop rate
   * decreases per interval. We only step up the global drop rate when the average cluster latency
   * is higher than the highWaterMark, and only step down the global drop rate when the average
   * cluster latency is lower than the global drop rate.
   *
   * This code is thread reentrant. Multiple threads can potentially call this concurrently, and so
   * callers must pass in the DegraderLoadBalancerState that they based their shouldUpdate() call on.
   * The multiple threads may have different views of the trackerClients latency, but this is
   * ok as the new state in the end will have only taken one action (either loadbalance or
   * call-dropping with at most one step). Currently we will not call this concurrently, as
   * checkUpdateState will control entry to a single thread.
   *
   * @param clusterGenerationId
   * @param trackerClients
   * @param oldState
   * @param config
   */
private static DegraderLoadBalancerState updateState(long clusterGenerationId, List<TrackerClient> trackerClients, DegraderLoadBalancerState oldState, DegraderLoadBalancerStrategyConfig config) {
    debug(_log, "updating state for: ", trackerClients);
    double sumOfClusterLatencies = 0.0;
    double computedClusterDropSum = 0.0;
    double computedClusterWeight = 0.0;
    long totalClusterCallCount = 0;
    boolean hashRingChanges = false;
    boolean recoveryMapChanges = false;
    DegraderLoadBalancerState.Strategy strategy = oldState.getStrategy();
    Map<TrackerClient, Double> oldRecoveryMap = oldState.getRecoveryMap();
    Map<TrackerClient, Double> newRecoveryMap = new HashMap<TrackerClient, Double>(oldRecoveryMap);
    double currentOverrideDropRate = oldState.getCurrentOverrideDropRate();
    double initialRecoveryLevel = config.getInitialRecoveryLevel();
    double ringRampFactor = config.getRingRampFactor();
    int pointsPerWeight = config.getPointsPerWeight();
    DegraderLoadBalancerState newState;
    for (TrackerClient client : trackerClients) {
        double averageLatency = client.getDegraderControl(DEFAULT_PARTITION_ID).getLatency();
        long callCount = client.getDegraderControl(DEFAULT_PARTITION_ID).getCallCount();
        oldState.getPreviousMaxDropRate().put(client, client.getDegraderControl(DEFAULT_PARTITION_ID).getMaxDropRate());
        sumOfClusterLatencies += averageLatency * callCount;
        totalClusterCallCount += callCount;
        double clientDropRate = client.getDegraderControl(DEFAULT_PARTITION_ID).getCurrentComputedDropRate();
        computedClusterDropSum += client.getPartitionWeight(DEFAULT_PARTITION_ID) * clientDropRate;
        computedClusterWeight += client.getPartitionWeight(DEFAULT_PARTITION_ID);
        boolean recoveryMapContainsClient = newRecoveryMap.containsKey(client);
        // points in the hash ring for the clients.
        if (callCount == 0) {
            // due solely to low volume.
            if (recoveryMapContainsClient) {
                // it may do nothing.
                if (strategy == DegraderLoadBalancerState.Strategy.LOAD_BALANCE) {
                    double oldMaxDropRate = client.getDegraderControl(DEFAULT_PARTITION_ID).getMaxDropRate();
                    double transmissionRate = 1.0 - oldMaxDropRate;
                    if (transmissionRate <= 0.0) {
                        // We use the initialRecoveryLevel to indicate how many points to initially set
                        // the tracker client to when traffic has stopped flowing to this node.
                        transmissionRate = initialRecoveryLevel;
                    } else {
                        transmissionRate *= ringRampFactor;
                        transmissionRate = Math.min(transmissionRate, 1.0);
                    }
                    double newMaxDropRate = 1.0 - transmissionRate;
                    client.getDegraderControl(DEFAULT_PARTITION_ID).setMaxDropRate(newMaxDropRate);
                }
                recoveryMapChanges = true;
            }
        } else //else we don't really need to change the client maxDropRate.
        if (recoveryMapContainsClient) {
            // else if the recovery map contains the client and the call count was > 0
            // tough love here, once the rehab clients start taking traffic, we
            // restore their maxDropRate to it's original value, and unenroll them
            // from the program.
            // This is safe because the hash ring points are controlled by the
            // computedDropRate variable, and the call dropping rate is controlled by
            // the overrideDropRate. The maxDropRate only serves to cap the computedDropRate and
            // overrideDropRate.
            // We store the maxDropRate and restore it here because the initialRecoveryLevel could
            // potentially be higher than what the default maxDropRate allowed. (the maxDropRate doesn't
            // necessarily have to be 1.0). For instance, if the maxDropRate was 0.99, and the
            // initialRecoveryLevel was 0.05  then we need to store the old maxDropRate.
            client.getDegraderControl(DEFAULT_PARTITION_ID).setMaxDropRate(newRecoveryMap.get(client));
            newRecoveryMap.remove(client);
            recoveryMapChanges = true;
        }
    }
    double computedClusterDropRate = computedClusterDropSum / computedClusterWeight;
    debug(_log, "total cluster call count: ", totalClusterCallCount);
    debug(_log, "computed cluster drop rate for ", trackerClients.size(), " nodes: ", computedClusterDropRate);
    if (oldState.getClusterGenerationId() == clusterGenerationId && totalClusterCallCount <= 0 && !recoveryMapChanges) {
        // if the cluster has not been called recently (total cluster call count is <= 0)
        // and we already have a state with the same set of URIs (same cluster generation),
        // and no clients are in rehab, then don't change anything.
        debug(_log, "New state is the same as the old state so we're not changing anything. Old state = ", oldState, ", config=", config);
        return new DegraderLoadBalancerState(oldState, clusterGenerationId, config.getUpdateIntervalMs(), config.getClock().currentTimeMillis());
    }
    // update our overrides.
    double newCurrentAvgClusterLatency = -1;
    if (totalClusterCallCount > 0) {
        newCurrentAvgClusterLatency = sumOfClusterLatencies / totalClusterCallCount;
    }
    debug(_log, "average cluster latency: ", newCurrentAvgClusterLatency);
    // This points map stores how many hash map points to allocate for each tracker client.
    Map<URI, Integer> points = new HashMap<URI, Integer>();
    Map<URI, Integer> oldPointsMap = oldState.getPointsMap();
    for (TrackerClient client : trackerClients) {
        double successfulTransmissionWeight;
        URI clientUri = client.getUri();
        // Don't take into account cluster health when calculating the number of points
        // for each client. This is because the individual clients already take into account
        // latency, and a successfulTransmissionWeight can and should be made
        // independent of other nodes in the cluster. Otherwise, one unhealthy client in a small
        // cluster can take down the entire cluster if the avg latency is too high.
        // The global drop rate will take into account the cluster latency. High cluster-wide error
        // rates are not something d2 can address.
        //
        // this client's maxDropRate and currentComputedDropRate may have been adjusted if it's in the
        // rehab program (to gradually send traffic it's way).
        double dropRate = Math.min(client.getDegraderControl(DEFAULT_PARTITION_ID).getCurrentComputedDropRate(), client.getDegraderControl(DEFAULT_PARTITION_ID).getMaxDropRate());
        // calculate the weight as the probability of successful transmission to this
        // node divided by the probability of successful transmission to the entire
        // cluster
        successfulTransmissionWeight = client.getPartitionWeight(DEFAULT_PARTITION_ID) * (1.0 - dropRate);
        // calculate the weight as the probability of a successful transmission to this node
        // multiplied by the client's self-defined weight. thus, the node's final weight
        // takes into account both the self defined weight (to account for different
        // hardware in the same cluster) and the performance of the node (as defined by the
        // node's degrader).
        debug(_log, "computed new weight for uri ", clientUri, ": ", successfulTransmissionWeight);
        // keep track if we're making actual changes to the Hash Ring in this updateState.
        int newPoints = (int) (successfulTransmissionWeight * pointsPerWeight);
        if (newPoints == 0) {
            // We are choking off traffic to this tracker client.
            // Enroll this tracker client in the recovery program so that
            // we can make sure it still gets some traffic
            Double oldMaxDropRate = client.getDegraderControl(DEFAULT_PARTITION_ID).getMaxDropRate();
            // set the default recovery level.
            newPoints = (int) (initialRecoveryLevel * pointsPerWeight);
            // Keep track of the original maxDropRate
            if (!newRecoveryMap.containsKey(client)) {
                // keep track of this client,
                newRecoveryMap.put(client, oldMaxDropRate);
                client.getDegraderControl(DEFAULT_PARTITION_ID).setMaxDropRate(1.0 - initialRecoveryLevel);
            }
        }
        points.put(clientUri, newPoints);
        if (!oldPointsMap.containsKey(clientUri) || oldPointsMap.get(clientUri) != newPoints) {
            hashRingChanges = true;
        }
    }
    // if there were changes to the members of the cluster
    if ((strategy == DegraderLoadBalancerState.Strategy.LOAD_BALANCE && hashRingChanges == true) || // strategy
    oldState.getClusterGenerationId() != clusterGenerationId) {
        // atomic overwrite
        // try Call Dropping next time we updateState.
        newState = new DegraderLoadBalancerState(config.getUpdateIntervalMs(), clusterGenerationId, points, config.getClock().currentTimeMillis(), DegraderLoadBalancerState.Strategy.CALL_DROPPING, currentOverrideDropRate, newCurrentAvgClusterLatency, true, newRecoveryMap, oldState.getServiceName(), oldState.getDegraderProperties(), totalClusterCallCount);
        logState(oldState, newState, config, trackerClients);
    } else {
        // time to try call dropping strategy, if necessary.
        // we are explicitly setting the override drop rate to a number between 0 and 1, inclusive.
        double newDropLevel = Math.max(0.0, currentOverrideDropRate);
        // to get the cluster latency stabilized
        if (newCurrentAvgClusterLatency > 0 && totalClusterCallCount >= config.getMinClusterCallCountHighWaterMark()) {
            // statistically significant
            if (newCurrentAvgClusterLatency >= config.getHighWaterMark() && currentOverrideDropRate != 1.0) {
                // if the cluster latency is too high and we can drop more traffic
                newDropLevel = Math.min(1.0, newDropLevel + config.getGlobalStepUp());
            } else if (newCurrentAvgClusterLatency <= config.getLowWaterMark() && currentOverrideDropRate != 0.0) {
                // else if the cluster latency is good and we can reduce the override drop rate
                newDropLevel = Math.max(0.0, newDropLevel - config.getGlobalStepDown());
            }
        // else the averageClusterLatency is between Low and High, or we can't change anything more,
        // then do not change anything.
        } else if (newCurrentAvgClusterLatency > 0 && totalClusterCallCount >= config.getMinClusterCallCountLowWaterMark()) {
            //but we might recover a bit if the latency is healthy
            if (newCurrentAvgClusterLatency <= config.getLowWaterMark() && currentOverrideDropRate != 0.0) {
                // the cluster latency is good and we can reduce the override drop rate
                newDropLevel = Math.max(0.0, newDropLevel - config.getGlobalStepDown());
            }
        // else the averageClusterLatency is somewhat high but since the qps is not that high, we shouldn't degrade
        } else {
            // if we enter here that means we have very low traffic. We should reduce the overrideDropRate, if possible.
            // when we have below 1 QPS traffic, we should be pretty confident that the cluster can handle very low
            // traffic. Of course this is depending on the MinClusterCallCountLowWaterMark that the service owner sets.
            // Another possible cause for this is if we had somehow choked off all traffic to the cluster, most
            // likely in a one node/small cluster scenario. Obviously, we can't check latency here,
            // we'll have to rely on the metric in the next updateState. If the cluster is still having
            // latency problems, then we will oscillate between off and letting a little traffic through,
            // and that is acceptable. If the latency, though high, is deemed acceptable, then the
            // watermarks can be adjusted to let more traffic through.
            newDropLevel = Math.max(0.0, newDropLevel - config.getGlobalStepDown());
        }
        if (newDropLevel != currentOverrideDropRate) {
            overrideClusterDropRate(newDropLevel, trackerClients);
        }
        // don't change the points map or the recoveryMap, but try load balancing strategy next time.
        newState = new DegraderLoadBalancerState(config.getUpdateIntervalMs(), clusterGenerationId, oldPointsMap, config.getClock().currentTimeMillis(), DegraderLoadBalancerState.Strategy.LOAD_BALANCE, newDropLevel, newCurrentAvgClusterLatency, true, oldRecoveryMap, oldState.getServiceName(), oldState.getDegraderProperties(), totalClusterCallCount);
        logState(oldState, newState, config, trackerClients);
        points = oldPointsMap;
    }
    // adjust the min call count for each client based on the hash ring reduction and call dropping
    // fraction.
    overrideMinCallCount(currentOverrideDropRate, trackerClients, points, pointsPerWeight);
    return newState;
}

Example 87 with TrackerClient

use of com.linkedin.d2.balancer.clients.TrackerClient in project rest.li by linkedin.

the class DegraderLoadBalancerStrategyV2 method isNewStateHealthy.

static boolean isNewStateHealthy(DegraderLoadBalancerState newState, DegraderLoadBalancerStrategyConfig config, List<TrackerClient> trackerClients) {
    if (newState.getCurrentAvgClusterLatency() > config.getLowWaterMark()) {
        return false;
    }
    Map<URI, Integer> pointsMap = newState.getPointsMap();
    for (TrackerClient client : trackerClients) {
        int perfectHealth = (int) (client.getPartitionWeight(DEFAULT_PARTITION_ID) * config.getPointsPerWeight());
        Integer point = pointsMap.get(client.getUri());
        if (point < perfectHealth) {
            return false;
        }
    }
    return true;
}

Example 88 with TrackerClient

use of com.linkedin.d2.balancer.clients.TrackerClient in project rest.li by linkedin.

the class SimpleLoadBalancer method getPotentialClients.

private List<TrackerClient> getPotentialClients(String serviceName, ServiceProperties serviceProperties, Set<URI> possibleUris) {
    List<TrackerClient> clientsToLoadBalance = new ArrayList<TrackerClient>();
    if (possibleUris != null) {
        for (URI possibleUri : possibleUris) {
            // don't pay attention to this uri if it's banned
            if (!serviceProperties.isBanned(possibleUri)) {
                TrackerClient possibleTrackerClient = _state.getClient(serviceName, possibleUri);
                if (possibleTrackerClient != null) {
                    clientsToLoadBalance.add(possibleTrackerClient);
                }
            } else {
                warn(_log, "skipping banned uri: ", possibleUri);
            }
        }
    }
    debug(_log, "got clients to load balancer for ", serviceName, ": ", clientsToLoadBalance);
    return clientsToLoadBalance;
}

Example 89 with TrackerClient

use of com.linkedin.d2.balancer.clients.TrackerClient in project rest.li by linkedin.

the class SimpleLoadBalancerState method getTrackerClient.

private TrackerClient getTrackerClient(String serviceName, URI uri, Map<Integer, PartitionData> partitionDataMap, DegraderImpl.Config config, Clock clk, long callTrackerInterval, String errorStatusPattern) {
    Map<String, TransportClient> clientsByScheme = _serviceClients.get(serviceName);
    if (clientsByScheme == null) {
        _log.error("getTrackerClient: unknown service name {} for URI {} and partitionDataMap {}", new Object[] { serviceName, uri, partitionDataMap });
        return null;
    }
    TransportClient client = clientsByScheme.get(uri.getScheme().toLowerCase());
    if (client == null) {
        // logging this at debug because there may be situations where a service may want some of its
        // clients talking https while others are ok using http.
        debug(_log, "No TransportClient for scheme {}, service {}, URI {} and partitionDataMap {}. " + "This client may not be configured to handle URIs in this scheme.", new Object[] { uri.getScheme(), serviceName, uri, partitionDataMap });
        return null;
    }
    TrackerClient trackerClient = new TrackerClient(uri, partitionDataMap, client, clk, config, callTrackerInterval, errorStatusPattern);
    return trackerClient;
}

Example 90 with TrackerClient

use of com.linkedin.d2.balancer.clients.TrackerClient in project rest.li by linkedin.

the class DegraderLoadBalancerTest method testCallDroppingHelper.

private void testCallDroppingHelper(DegraderLoadBalancerStrategyAdapter strategyAdapter, List<TrackerClient> clients, TestClock clock, Long timeInterval) {
    //test clusterOverrideDropRate won't increase even though latency is 3000 ms because the traffic is low
    callClients(3000, 0.2, clients, clock, timeInterval, false, false);
    strategyAdapter.setStrategyToCallDrop();
    URIRequest request = new URIRequest(clients.get(0).getUri());
    getTrackerClient(strategyAdapter, request, new RequestContext(), 1, clients);
    assertTrue(isEqual(0.0d, strategyAdapter.getCurrentOverrideDropRate()));
    //if we increase the QPS from 0.2 to 25, then we'll start dropping calls
    callClients(3000, 25, clients, clock, timeInterval, false, false);
    strategyAdapter.setStrategyToCallDrop();
    getTrackerClient(strategyAdapter, request, new RequestContext(), 1, clients);
    assertTrue(isEqual(0.2d, strategyAdapter.getCurrentOverrideDropRate()));
    ;
    // if we set the QPS to be somewhere below high and low water mark then the drop rate stays the same
    // even though the latency is high
    callClients(3000, 2, clients, clock, timeInterval, false, false);
    strategyAdapter.setStrategyToCallDrop();
    getTrackerClient(strategyAdapter, request, new RequestContext(), 1, clients);
    assertTrue(isEqual(0.2d, strategyAdapter.getCurrentOverrideDropRate()));
    // now we want to degrade the cluster even further
    callClients(3000, 25, clients, clock, timeInterval, false, false);
    strategyAdapter.setStrategyToCallDrop();
    getTrackerClient(strategyAdapter, request, new RequestContext(), 1, clients);
    assertTrue(isEqual(0.4d, strategyAdapter.getCurrentOverrideDropRate()));
    callClients(3000, 25, clients, clock, timeInterval, false, false);
    strategyAdapter.setStrategyToCallDrop();
    getTrackerClient(strategyAdapter, request, new RequestContext(), 1, clients);
    assertTrue(isEqual(0.6d, strategyAdapter.getCurrentOverrideDropRate()));
    callClients(3000, 25, clients, clock, timeInterval, false, false);
    strategyAdapter.setStrategyToCallDrop();
    getTrackerClient(strategyAdapter, request, new RequestContext(), 1, clients);
    assertTrue(isEqual(0.8d, strategyAdapter.getCurrentOverrideDropRate()));
    callClients(3000, 25, clients, clock, timeInterval, false, false);
    strategyAdapter.setStrategyToCallDrop();
    getTrackerClient(strategyAdapter, request, new RequestContext(), 1, clients);
    assertTrue(isEqual(1.0d, strategyAdapter.getCurrentOverrideDropRate()));
    //if we have qps below lowWaterMark, we will reduce drop rate even though latency is high
    callClients(3000, 0.5, clients, clock, timeInterval, false, false);
    strategyAdapter.setStrategyToCallDrop();
    getTrackerClient(strategyAdapter, request, new RequestContext(), 1, clients);
    assertTrue(isEqual(0.8d, strategyAdapter.getCurrentOverrideDropRate()));
    //if we have qps below lowWaterMark and qps is low, we will also reduce drop rate
    callClients(100, 0.5, clients, clock, timeInterval, false, false);
    strategyAdapter.setStrategyToCallDrop();
    getTrackerClient(strategyAdapter, request, new RequestContext(), 1, clients);
    assertTrue(isEqual(0.6d, strategyAdapter.getCurrentOverrideDropRate()));
    //if we have qps between lowWaterMark and highWaterMark and latency is low, we will reduce drop rate
    callClients(100, 2, clients, clock, timeInterval, false, false);
    strategyAdapter.setStrategyToCallDrop();
    getTrackerClient(strategyAdapter, request, new RequestContext(), 1, clients);
    assertTrue(isEqual(0.4d, strategyAdapter.getCurrentOverrideDropRate()));
    //if we have qps higher than highWaterMark and latency is low, we will reduce drop rate
    callClients(100, 25, clients, clock, timeInterval, false, false);
    strategyAdapter.setStrategyToCallDrop();
    getTrackerClient(strategyAdapter, request, new RequestContext(), 1, clients);
    assertTrue(isEqual(0.2d, strategyAdapter.getCurrentOverrideDropRate()));
}