Search in sources :

Example 11 with Diagnosis

use of com.microsoft.dhalion.diagnoser.Diagnosis in project incubator-heron by apache.

the class DataSkewDiagnoser method diagnose.

@Override
public Diagnosis diagnose(List<Symptom> symptoms) {
    List<Symptom> bpSymptoms = getBackPressureSymptoms(symptoms);
    Map<String, ComponentMetrics> processingRateSkewComponents = getProcessingRateSkewComponents(symptoms);
    Map<String, ComponentMetrics> waitQDisparityComponents = getWaitQDisparityComponents(symptoms);
    if (bpSymptoms.isEmpty() || processingRateSkewComponents.isEmpty() || waitQDisparityComponents.isEmpty()) {
        // Since there is no back pressure or disparate execute count, no action is needed
        return null;
    } else if (bpSymptoms.size() > 1) {
        // TODO handle cases where multiple detectors create back pressure symptom
        throw new IllegalStateException("Multiple back-pressure symptoms case");
    }
    ComponentMetrics bpMetrics = bpSymptoms.iterator().next().getComponent();
    // verify data skew, larger queue size and back pressure for the same component exists
    ComponentMetrics exeCountMetrics = processingRateSkewComponents.get(bpMetrics.getName());
    ComponentMetrics pendingBufferMetrics = waitQDisparityComponents.get(bpMetrics.getName());
    if (exeCountMetrics == null || pendingBufferMetrics == null) {
        // for the component with back pressure. This is not a data skew case
        return null;
    }
    ComponentMetrics mergedData = ComponentMetrics.merge(bpMetrics, ComponentMetrics.merge(exeCountMetrics, pendingBufferMetrics));
    ComponentMetricsHelper compStats = new ComponentMetricsHelper(mergedData);
    compStats.computeBpStats();
    MetricsStats exeStats = compStats.computeMinMaxStats(METRIC_EXE_COUNT);
    MetricsStats bufferStats = compStats.computeMinMaxStats(METRIC_BUFFER_SIZE);
    Symptom resultSymptom = null;
    for (InstanceMetrics boltMetrics : compStats.getBoltsWithBackpressure()) {
        double exeCount = boltMetrics.getMetricValueSum(METRIC_EXE_COUNT.text());
        double bufferSize = boltMetrics.getMetricValueSum(METRIC_BUFFER_SIZE.text());
        double bpValue = boltMetrics.getMetricValueSum(METRIC_BACK_PRESSURE.text());
        if (exeStats.getMetricMax() < 1.10 * exeCount && bufferStats.getMetricMax() < 2 * bufferSize) {
            LOG.info(String.format("DataSkew: %s back-pressure(%s), high execution count: %s and " + "high buffer size %s", boltMetrics.getName(), bpValue, exeCount, bufferSize));
            resultSymptom = new Symptom(SYMPTOM_DATA_SKEW.text(), mergedData);
        }
    }
    return resultSymptom != null ? new Diagnosis(DIAGNOSIS_DATA_SKEW.text(), resultSymptom) : null;
}
Also used : InstanceMetrics(com.microsoft.dhalion.metrics.InstanceMetrics) ComponentMetricsHelper(com.twitter.heron.healthmgr.common.ComponentMetricsHelper) Diagnosis(com.microsoft.dhalion.diagnoser.Diagnosis) Symptom(com.microsoft.dhalion.detector.Symptom) ComponentMetrics(com.microsoft.dhalion.metrics.ComponentMetrics) MetricsStats(com.twitter.heron.healthmgr.common.MetricsStats)

Example 12 with Diagnosis

use of com.microsoft.dhalion.diagnoser.Diagnosis in project incubator-heron by apache.

the class UnderProvisioningDiagnoser method diagnose.

@Override
public Diagnosis diagnose(List<Symptom> symptoms) {
    List<Symptom> bpSymptoms = getBackPressureSymptoms(symptoms);
    Map<String, ComponentMetrics> processingRateSkewComponents = getProcessingRateSkewComponents(symptoms);
    Map<String, ComponentMetrics> waitQDisparityComponents = getWaitQDisparityComponents(symptoms);
    if (bpSymptoms.isEmpty() || !processingRateSkewComponents.isEmpty() || !waitQDisparityComponents.isEmpty()) {
        // and buffer sizes, no action is needed
        return null;
    } else if (bpSymptoms.size() > 1) {
        // TODO handle cases where multiple detectors create back pressure symptom
        throw new IllegalStateException("Multiple back-pressure symptoms case");
    }
    ComponentMetrics bpMetrics = bpSymptoms.iterator().next().getComponent();
    ComponentMetricsHelper compStats = new ComponentMetricsHelper(bpMetrics);
    compStats.computeBpStats();
    LOG.info(String.format("UNDER_PROVISIONING: %s back-pressure(%s) and similar processing rates " + "and buffer sizes", bpMetrics.getName(), compStats.getTotalBackpressure()));
    Symptom resultSymptom = new Symptom(SYMPTOM_UNDER_PROVISIONING.text(), bpMetrics);
    return new Diagnosis(DIAGNOSIS_UNDER_PROVISIONING.text(), resultSymptom);
}
Also used : ComponentMetricsHelper(com.twitter.heron.healthmgr.common.ComponentMetricsHelper) Diagnosis(com.microsoft.dhalion.diagnoser.Diagnosis) Symptom(com.microsoft.dhalion.detector.Symptom) ComponentMetrics(com.microsoft.dhalion.metrics.ComponentMetrics)

Example 13 with Diagnosis

use of com.microsoft.dhalion.diagnoser.Diagnosis in project incubator-heron by apache.

the class ScaleUpResolver method resolve.

@Override
public List<Action> resolve(List<Diagnosis> diagnosis) {
    for (Diagnosis diagnoses : diagnosis) {
        Symptom bpSymptom = diagnoses.getSymptoms().get(SYMPTOM_UNDER_PROVISIONING.text());
        if (bpSymptom == null || bpSymptom.getComponents().isEmpty()) {
            // nothing to fix as there is no back pressure
            continue;
        }
        if (bpSymptom.getComponents().size() > 1) {
            throw new UnsupportedOperationException("Multiple components with back pressure symptom");
        }
        ComponentMetrics bpComponent = bpSymptom.getComponent();
        int newParallelism = computeScaleUpFactor(bpComponent);
        Map<String, Integer> changeRequest = new HashMap<>();
        changeRequest.put(bpComponent.getName(), newParallelism);
        PackingPlan currentPackingPlan = packingPlanProvider.get();
        PackingPlan newPlan = buildNewPackingPlan(changeRequest, currentPackingPlan);
        if (newPlan == null) {
            return null;
        }
        Scheduler.UpdateTopologyRequest updateTopologyRequest = Scheduler.UpdateTopologyRequest.newBuilder().setCurrentPackingPlan(getSerializedPlan(currentPackingPlan)).setProposedPackingPlan(getSerializedPlan(newPlan)).build();
        LOG.info("Sending Updating topology request: " + updateTopologyRequest);
        if (!schedulerClient.updateTopology(updateTopologyRequest)) {
            throw new RuntimeException(String.format("Failed to update topology with Scheduler, " + "updateTopologyRequest=%s", updateTopologyRequest));
        }
        TopologyUpdate action = new TopologyUpdate();
        LOG.info("Broadcasting topology update event");
        eventManager.onEvent(action);
        LOG.info("Scheduler updated topology successfully.");
        List<Action> actions = new ArrayList<>();
        actions.add(action);
        return actions;
    }
    return null;
}
Also used : Action(com.microsoft.dhalion.resolver.Action) HashMap(java.util.HashMap) Scheduler(com.twitter.heron.proto.scheduler.Scheduler) PackingPlan(com.twitter.heron.spi.packing.PackingPlan) ArrayList(java.util.ArrayList) TopologyUpdate(com.twitter.heron.healthmgr.common.HealthManagerEvents.TopologyUpdate) Diagnosis(com.microsoft.dhalion.diagnoser.Diagnosis) Symptom(com.microsoft.dhalion.detector.Symptom) ComponentMetrics(com.microsoft.dhalion.metrics.ComponentMetrics)

Example 14 with Diagnosis

use of com.microsoft.dhalion.diagnoser.Diagnosis in project incubator-heron by apache.

the class DataSkewDiagnoserTest method failsIfNoDataSkewSymptom.

@Test
public void failsIfNoDataSkewSymptom() {
    List<Symptom> symptoms = TestUtils.createBpSymptomList(123);
    Diagnosis result = new DataSkewDiagnoser().diagnose(symptoms);
    assertNull(result);
}
Also used : Diagnosis(com.microsoft.dhalion.diagnoser.Diagnosis) Symptom(com.microsoft.dhalion.detector.Symptom) Test(org.junit.Test)

Example 15 with Diagnosis

use of com.microsoft.dhalion.diagnoser.Diagnosis in project incubator-heron by apache.

the class UnderProvisioningDiagnoserTest method diagnosisWhen1Of1InstanceInBP.

@Test
public void diagnosisWhen1Of1InstanceInBP() {
    List<Symptom> symptoms = TestUtils.createBpSymptomList(123, 0);
    // symptoms.add(TestUtils.createLargeWaitQSymptom(5000));
    Diagnosis result = new UnderProvisioningDiagnoser().diagnose(symptoms);
    validateDiagnosis(result);
}
Also used : Diagnosis(com.microsoft.dhalion.diagnoser.Diagnosis) Symptom(com.microsoft.dhalion.detector.Symptom) Test(org.junit.Test)

Aggregations

Diagnosis (com.microsoft.dhalion.diagnoser.Diagnosis)16 Symptom (com.microsoft.dhalion.detector.Symptom)15 Test (org.junit.Test)11 ComponentMetrics (com.microsoft.dhalion.metrics.ComponentMetrics)6 InstanceMetrics (com.microsoft.dhalion.metrics.InstanceMetrics)3 Action (com.microsoft.dhalion.resolver.Action)3 ComponentMetricsHelper (com.twitter.heron.healthmgr.common.ComponentMetricsHelper)3 ArrayList (java.util.ArrayList)3 MetricsStats (com.twitter.heron.healthmgr.common.MetricsStats)2 PackingPlan (com.twitter.heron.spi.packing.PackingPlan)2 HashMap (java.util.HashMap)2 TopologyAPI (com.twitter.heron.api.generated.TopologyAPI)1 ContainerRestart (com.twitter.heron.healthmgr.common.HealthManagerEvents.ContainerRestart)1 TopologyUpdate (com.twitter.heron.healthmgr.common.HealthManagerEvents.TopologyUpdate)1 PackingPlanProvider (com.twitter.heron.healthmgr.common.PackingPlanProvider)1 Scheduler (com.twitter.heron.proto.scheduler.Scheduler)1 UpdateTopologyRequest (com.twitter.heron.proto.scheduler.Scheduler.UpdateTopologyRequest)1 ISchedulerClient (com.twitter.heron.scheduler.client.ISchedulerClient)1 Config (com.twitter.heron.spi.common.Config)1