use of com.twitter.heron.healthmgr.common.ComponentMetricsHelper in project incubator-heron by apache.
the class GrowingWaitQueueDetector method detect.
/**
* Detects all components unable to keep up with input load, hence having a growing pending buffer
* or wait queue
*
* @return A collection of all components executing slower than input rate.
*/
@Override
public List<Symptom> detect() {
ArrayList<Symptom> result = new ArrayList<>();
Map<String, ComponentMetrics> bufferSizes = pendingBufferSensor.get();
for (ComponentMetrics compMetrics : bufferSizes.values()) {
ComponentMetricsHelper compStats = new ComponentMetricsHelper(compMetrics);
compStats.computeBufferSizeTrend();
if (compStats.getMaxBufferChangeRate() > rateLimit) {
LOG.info(String.format("Detected growing wait queues for %s, max rate %f", compMetrics.getName(), compStats.getMaxBufferChangeRate()));
result.add(new Symptom(SYMPTOM_GROWING_WAIT_Q.text(), compMetrics));
}
}
return result;
}
use of com.twitter.heron.healthmgr.common.ComponentMetricsHelper in project incubator-heron by apache.
the class SlowInstanceDiagnoser 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() || waitQDisparityComponents.isEmpty() || !processingRateSkewComponents.isEmpty()) {
// execution 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 wait Q disparity and back pressure for the same component exists
ComponentMetrics pendingBufferMetrics = waitQDisparityComponents.get(bpMetrics.getName());
if (pendingBufferMetrics == null) {
// no wait Q disparity for the component with back pressure. There is no slow instance
return null;
}
ComponentMetrics mergedData = ComponentMetrics.merge(bpMetrics, pendingBufferMetrics);
ComponentMetricsHelper compStats = new ComponentMetricsHelper(mergedData);
compStats.computeBpStats();
MetricsStats bufferStats = compStats.computeMinMaxStats(METRIC_BUFFER_SIZE);
Symptom resultSymptom = null;
for (InstanceMetrics boltMetrics : compStats.getBoltsWithBackpressure()) {
double bufferSize = boltMetrics.getMetricValueSum(METRIC_BUFFER_SIZE.text());
double bpValue = boltMetrics.getMetricValueSum(METRIC_BACK_PRESSURE.text());
if (bufferStats.getMetricMax() < bufferSize * 2) {
LOG.info(String.format("SLOW: %s back-pressure(%s) and high buffer size: %s " + "and similar processing rates", boltMetrics.getName(), bpValue, bufferSize));
resultSymptom = new Symptom(SYMPTOM_SLOW_INSTANCE.text(), mergedData);
}
}
return resultSymptom != null ? new Diagnosis(DIAGNOSIS_SLOW_INSTANCE.text(), resultSymptom) : null;
}
use of com.twitter.heron.healthmgr.common.ComponentMetricsHelper in project incubator-heron by apache.
the class SkewDetector method detect.
/**
* Detects components experiencing data skew, instances with vastly different execute counts.
*
* @return A collection of affected components
*/
@Override
public List<Symptom> detect() {
ArrayList<Symptom> result = new ArrayList<>();
Map<String, ComponentMetrics> metrics = sensor.get();
for (ComponentMetrics compMetrics : metrics.values()) {
ComponentMetricsHelper compStats = new ComponentMetricsHelper(compMetrics);
MetricsStats stats = compStats.computeMinMaxStats(sensor.getMetricName());
if (stats.getMetricMax() > skewRatio * stats.getMetricMin()) {
LOG.info(String.format("Detected skew for %s, min = %f, max = %f", compMetrics.getName(), stats.getMetricMin(), stats.getMetricMax()));
result.add(new Symptom(symptomName.text(), compMetrics));
}
}
return result;
}
use of com.twitter.heron.healthmgr.common.ComponentMetricsHelper in project incubator-heron by apache.
the class LargeWaitQueueDetector method detect.
/**
* Detects all components unable to keep up with input load, hence having a large pending buffer
* or wait queue
*
* @return A collection of all components executing slower than input rate.
*/
@Override
public List<Symptom> detect() {
ArrayList<Symptom> result = new ArrayList<>();
Map<String, ComponentMetrics> bufferSizes = pendingBufferSensor.get();
for (ComponentMetrics compMetrics : bufferSizes.values()) {
ComponentMetricsHelper compStats = new ComponentMetricsHelper(compMetrics);
MetricsStats stats = compStats.computeMinMaxStats(METRIC_BUFFER_SIZE.text());
if (stats.getMetricMin() > sizeLimit) {
LOG.info(String.format("Detected large wait queues for %s, smallest queue is %f", compMetrics.getName(), stats.getMetricMin()));
result.add(new Symptom(SYMPTOM_LARGE_WAIT_Q.text(), compMetrics));
}
}
return result;
}
use of com.twitter.heron.healthmgr.common.ComponentMetricsHelper in project incubator-heron by apache.
the class BackPressureDetector method detect.
/**
* Detects all components initiating backpressure above the configured limit. Normally there
* will be only one component
*
* @return A collection of all components causing backpressure.
*/
@Override
public List<Symptom> detect() {
ArrayList<Symptom> result = new ArrayList<>();
Map<String, ComponentMetrics> backpressureMetrics = bpSensor.get();
for (ComponentMetrics compMetrics : backpressureMetrics.values()) {
ComponentMetricsHelper compStats = new ComponentMetricsHelper(compMetrics);
compStats.computeBpStats();
if (compStats.getTotalBackpressure() > noiseFilterMillis) {
LOG.info(String.format("Detected back pressure for %s, total back pressure is %f", compMetrics.getName(), compStats.getTotalBackpressure()));
result.add(new Symptom(SYMPTOM_BACK_PRESSURE.text(), compMetrics));
}
}
return result;
}
Aggregations