use of com.datatorrent.stram.plan.physical.OperatorStatus.PortStatus in project apex-core by apache.
the class StreamingContainerManager method fillPhysicalOperatorInfo.
private OperatorInfo fillPhysicalOperatorInfo(PTOperator operator) {
OperatorInfo oi = new OperatorInfo();
oi.container = operator.getContainer().getExternalId();
oi.host = operator.getContainer().host;
oi.id = Integer.toString(operator.getId());
oi.name = operator.getName();
oi.className = operator.getOperatorMeta().getOperator().getClass().getName();
oi.status = operator.getState().toString();
if (operator.isUnifier()) {
oi.unifierClass = operator.getUnifierClass().getName();
}
oi.logicalName = operator.getOperatorMeta().getName();
OperatorStatus os = operator.stats;
oi.recordingId = os.recordingId;
oi.totalTuplesProcessed = os.totalTuplesProcessed.get();
oi.totalTuplesEmitted = os.totalTuplesEmitted.get();
oi.tuplesProcessedPSMA = os.tuplesProcessedPSMA.get();
oi.tuplesEmittedPSMA = os.tuplesEmittedPSMA.get();
oi.cpuPercentageMA = os.cpuNanosPMSMA.getAvg() / 10000;
oi.latencyMA = os.latencyMA.getAvg();
oi.failureCount = operator.failureCount;
oi.recoveryWindowId = toWsWindowId(operator.getRecoveryCheckpoint().windowId);
oi.currentWindowId = toWsWindowId(os.currentWindowId.get());
if (os.lastHeartbeat != null) {
oi.lastHeartbeat = os.lastHeartbeat.getGeneratedTms();
}
if (os.checkpointStats != null) {
oi.checkpointTime = os.checkpointStats.checkpointTime;
oi.checkpointStartTime = os.checkpointStats.checkpointStartTime;
}
oi.checkpointTimeMA = os.checkpointTimeMA.getAvg();
for (PortStatus ps : os.inputPortStatusList.values()) {
PortInfo pinfo = new PortInfo();
pinfo.name = ps.portName;
pinfo.type = "input";
pinfo.totalTuples = ps.totalTuples;
pinfo.tuplesPSMA = Math.round(ps.tuplesPMSMA.getAvg() * 1000);
pinfo.bufferServerBytesPSMA = Math.round(ps.bufferServerBytesPMSMA.getAvg() * 1000);
pinfo.queueSizeMA = ps.queueSizeMA.getAvg();
pinfo.recordingId = ps.recordingId;
oi.addPort(pinfo);
}
for (PortStatus ps : os.outputPortStatusList.values()) {
PortInfo pinfo = new PortInfo();
pinfo.name = ps.portName;
pinfo.type = "output";
pinfo.totalTuples = ps.totalTuples;
pinfo.tuplesPSMA = Math.round(ps.tuplesPMSMA.getAvg() * 1000);
pinfo.bufferServerBytesPSMA = Math.round(ps.bufferServerBytesPMSMA.getAvg() * 1000);
pinfo.recordingId = ps.recordingId;
oi.addPort(pinfo);
}
oi.counters = os.getLastWindowedStats().size() > 0 ? os.getLastWindowedStats().get(os.getLastWindowedStats().size() - 1).counters : null;
oi.metrics = os.getLastWindowedStats().size() > 0 ? os.getLastWindowedStats().get(os.getLastWindowedStats().size() - 1).metrics : null;
return oi;
}
use of com.datatorrent.stram.plan.physical.OperatorStatus.PortStatus in project apex-core by apache.
the class StreamingContainerManagerTest method testProcessHeartbeat.
@Test
public void testProcessHeartbeat() throws Exception {
TestGeneratorInputOperator o1 = dag.addOperator("o1", TestGeneratorInputOperator.class);
dag.setOperatorAttribute(o1, OperatorContext.STATS_LISTENERS, Arrays.asList(new StatsListener[] { new PartitioningTest.PartitionLoadWatch() }));
dag.setAttribute(OperatorContext.STORAGE_AGENT, new MemoryStorageAgent());
StreamingContainerManager scm = new StreamingContainerManager(dag);
PhysicalPlan plan = scm.getPhysicalPlan();
Assert.assertEquals("number required containers", 1, plan.getContainers().size());
PTOperator o1p1 = plan.getOperators(dag.getMeta(o1)).get(0);
// assign container
String containerId = "container1";
StreamingContainerAgent sca = scm.assignContainer(new ContainerResource(0, containerId, "localhost", 512, 0, null), InetSocketAddress.createUnresolved("localhost", 0));
Assert.assertNotNull(sca);
Assert.assertEquals(PTContainer.State.ALLOCATED, o1p1.getContainer().getState());
Assert.assertEquals(PTOperator.State.PENDING_DEPLOY, o1p1.getState());
ContainerStats cstats = new ContainerStats(containerId);
ContainerHeartbeat hb = new ContainerHeartbeat();
hb.setContainerStats(cstats);
// get deploy request
ContainerHeartbeatResponse chr = scm.processHeartbeat(hb);
Assert.assertNotNull(chr.deployRequest);
Assert.assertEquals("" + chr.deployRequest, 1, chr.deployRequest.size());
Assert.assertEquals(PTContainer.State.ACTIVE, o1p1.getContainer().getState());
Assert.assertEquals("state " + o1p1, PTOperator.State.PENDING_DEPLOY, o1p1.getState());
// first operator heartbeat
OperatorHeartbeat ohb = new OperatorHeartbeat();
ohb.setNodeId(o1p1.getId());
ohb.setState(OperatorHeartbeat.DeployState.ACTIVE);
OperatorStats stats = new OperatorStats();
stats.checkpoint = new Checkpoint(2, 0, 0);
stats.windowId = 3;
stats.outputPorts = Lists.newArrayList();
PortStats ps = new PortStats(TestGeneratorInputOperator.OUTPUT_PORT);
ps.bufferServerBytes = 101;
ps.tupleCount = 1;
stats.outputPorts.add(ps);
ohb.windowStats = Lists.newArrayList(stats);
cstats.operators.add(ohb);
// activate operator
scm.processHeartbeat(hb);
Assert.assertEquals(PTContainer.State.ACTIVE, o1p1.getContainer().getState());
Assert.assertEquals("state " + o1p1, PTOperator.State.ACTIVE, o1p1.getState());
Assert.assertEquals("tuples " + o1p1, 1, o1p1.stats.totalTuplesEmitted.get());
Assert.assertEquals("tuples " + o1p1, 0, o1p1.stats.totalTuplesProcessed.get());
Assert.assertEquals("window " + o1p1, 3, o1p1.stats.currentWindowId.get());
Assert.assertEquals("port stats", 1, o1p1.stats.outputPortStatusList.size());
PortStatus o1p1ps = o1p1.stats.outputPortStatusList.get(TestGeneratorInputOperator.OUTPUT_PORT);
Assert.assertNotNull("port stats", o1p1ps);
Assert.assertEquals("port stats", 1, o1p1ps.totalTuples);
// second operator heartbeat
stats = new OperatorStats();
stats.checkpoint = new Checkpoint(2, 0, 0);
stats.windowId = 4;
stats.outputPorts = Lists.newArrayList();
ps = new PortStats(TestGeneratorInputOperator.OUTPUT_PORT);
ps.bufferServerBytes = 1;
ps.tupleCount = 1;
stats.outputPorts.add(ps);
ohb.windowStats = Lists.newArrayList(stats);
cstats.operators.clear();
cstats.operators.add(ohb);
scm.processHeartbeat(hb);
Assert.assertEquals("tuples " + o1p1, 2, o1p1.stats.totalTuplesEmitted.get());
Assert.assertEquals("window " + o1p1, 4, o1p1.stats.currentWindowId.get());
Assert.assertEquals("statsQueue " + o1p1, 2, o1p1.stats.listenerStats.size());
scm.processEvents();
Assert.assertEquals("statsQueue " + o1p1, 0, o1p1.stats.listenerStats.size());
Assert.assertEquals("lastStats " + o1p1, 2, o1p1.stats.lastWindowedStats.size());
}
use of com.datatorrent.stram.plan.physical.OperatorStatus.PortStatus in project apex-core by apache.
the class StreamingContainerManager method processHeartbeat.
/**
* process the heartbeat from each container.
* called by the RPC thread for each container. (i.e. called by multiple threads)
*
* @param heartbeat
* @return heartbeat response
*/
@SuppressWarnings("StatementWithEmptyBody")
public ContainerHeartbeatResponse processHeartbeat(ContainerHeartbeat heartbeat) {
long currentTimeMillis = clock.getTime();
final StreamingContainerAgent sca = this.containers.get(heartbeat.getContainerId());
if (sca == null || sca.container.getState() == PTContainer.State.KILLED) {
// could be orphaned container that was replaced and needs to terminate
LOG.error("Unknown container {}", heartbeat.getContainerId());
ContainerHeartbeatResponse response = new ContainerHeartbeatResponse();
response.shutdown = ShutdownType.ABORT;
return response;
}
// LOG.debug("{} {} {}", new Object[]{sca.container.containerId, sca.container.bufferServerAddress, sca.container.getState()});
if (sca.container.getState() == PTContainer.State.ALLOCATED) {
// capture dynamically assigned address from container
if (sca.container.bufferServerAddress == null && heartbeat.bufferServerHost != null) {
sca.container.bufferServerAddress = InetSocketAddress.createUnresolved(heartbeat.bufferServerHost, heartbeat.bufferServerPort);
LOG.info("Container {} buffer server: {}", sca.container.getExternalId(), sca.container.bufferServerAddress);
}
final long containerStartTime = System.currentTimeMillis();
sca.container.setState(PTContainer.State.ACTIVE);
sca.container.setStartedTime(containerStartTime);
sca.container.setFinishedTime(-1);
sca.jvmName = heartbeat.jvmName;
poolExecutor.submit(new Runnable() {
@Override
public void run() {
try {
containerFile.append(sca.getContainerInfo());
} catch (IOException ex) {
LOG.warn("Cannot write to container file");
}
for (PTOperator ptOp : sca.container.getOperators()) {
try {
JSONObject operatorInfo = new JSONObject();
operatorInfo.put("name", ptOp.getName());
operatorInfo.put("id", ptOp.getId());
operatorInfo.put("container", sca.container.getExternalId());
operatorInfo.put("startTime", containerStartTime);
operatorFile.append(operatorInfo);
} catch (IOException | JSONException ex) {
LOG.warn("Cannot write to operator file: ", ex);
}
}
}
});
}
sca.containerStackTrace = heartbeat.stackTrace;
if (heartbeat.restartRequested) {
LOG.error("Container {} restart request", sca.container.getExternalId());
containerStopRequests.put(sca.container.getExternalId(), sca.container.getExternalId());
}
sca.memoryMBFree = heartbeat.memoryMBFree;
sca.gcCollectionCount = heartbeat.gcCollectionCount;
sca.gcCollectionTime = heartbeat.gcCollectionTime;
sca.undeployOpers.clear();
sca.deployOpers.clear();
if (!this.deployChangeInProgress.get()) {
sca.deployCnt = this.deployChangeCnt;
}
Set<Integer> reportedOperators = Sets.newHashSetWithExpectedSize(sca.container.getOperators().size());
for (OperatorHeartbeat shb : heartbeat.getContainerStats().operators) {
long maxEndWindowTimestamp = 0;
reportedOperators.add(shb.nodeId);
PTOperator oper = this.plan.getAllOperators().get(shb.getNodeId());
if (oper == null) {
LOG.info("Heartbeat for unknown operator {} (container {})", shb.getNodeId(), heartbeat.getContainerId());
sca.undeployOpers.add(shb.nodeId);
continue;
}
if (shb.requestResponse != null) {
for (StatsListener.OperatorResponse obj : shb.requestResponse) {
if (obj instanceof OperatorResponse) {
// This is to identify platform requests
commandResponse.put((Long) obj.getResponseId(), obj.getResponse());
LOG.debug(" Got back the response {} for the request {}", obj, obj.getResponseId());
} else {
// This is to identify user requests
oper.stats.responses.add(obj);
}
}
}
// LOG.debug("heartbeat {} {}/{} {}", oper, oper.getState(), shb.getState(), oper.getContainer().getExternalId());
if (!(oper.getState() == PTOperator.State.ACTIVE && shb.getState() == OperatorHeartbeat.DeployState.ACTIVE)) {
// deploy state may require synchronization
processOperatorDeployStatus(oper, shb, sca);
}
oper.stats.lastHeartbeat = shb;
List<ContainerStats.OperatorStats> statsList = shb.getOperatorStatsContainer();
if (!statsList.isEmpty()) {
long tuplesProcessed = 0;
long tuplesEmitted = 0;
long totalCpuTimeUsed = 0;
int statCount = 0;
long maxDequeueTimestamp = -1;
oper.stats.recordingId = null;
final OperatorStatus status = oper.stats;
status.statsRevs.checkout();
for (Map.Entry<String, PortStatus> entry : status.inputPortStatusList.entrySet()) {
entry.getValue().recordingId = null;
}
for (Map.Entry<String, PortStatus> entry : status.outputPortStatusList.entrySet()) {
entry.getValue().recordingId = null;
}
for (ContainerStats.OperatorStats stats : statsList) {
if (stats == null) {
LOG.warn("Operator {} statistics list contains null element", shb.getNodeId());
continue;
}
/* report checkpoint-ed WindowId status of the operator */
if (stats.checkpoint instanceof Checkpoint) {
if (oper.getRecentCheckpoint() == null || oper.getRecentCheckpoint().windowId < stats.checkpoint.getWindowId()) {
addCheckpoint(oper, (Checkpoint) stats.checkpoint);
if (stats.checkpointStats != null) {
status.checkpointStats = stats.checkpointStats;
status.checkpointTimeMA.add(stats.checkpointStats.checkpointTime);
}
oper.failureCount = 0;
}
}
oper.stats.recordingId = stats.recordingId;
/* report all the other stuff */
// calculate the stats related to end window
// end window stats for a particular window id for a particular node
EndWindowStats endWindowStats = new EndWindowStats();
Collection<ContainerStats.OperatorStats.PortStats> ports = stats.inputPorts;
if (ports != null) {
Set<String> currentInputPortSet = Sets.newHashSetWithExpectedSize(ports.size());
for (ContainerStats.OperatorStats.PortStats s : ports) {
currentInputPortSet.add(s.id);
PortStatus ps = status.inputPortStatusList.get(s.id);
if (ps == null) {
ps = status.new PortStatus();
ps.portName = s.id;
status.inputPortStatusList.put(s.id, ps);
}
ps.totalTuples += s.tupleCount;
ps.recordingId = s.recordingId;
tuplesProcessed += s.tupleCount;
endWindowStats.dequeueTimestamps.put(s.id, s.endWindowTimestamp);
Pair<Integer, String> operatorPortName = new Pair<>(oper.getId(), s.id);
Long lastEndWindowTimestamp = operatorPortLastEndWindowTimestamps.get(operatorPortName);
if (lastEndWindowTimestamp == null) {
lastEndWindowTimestamp = lastStatsTimestamp;
}
long portElapsedMillis = Math.max(s.endWindowTimestamp - lastEndWindowTimestamp, 0);
// LOG.debug("=== PROCESSED TUPLE COUNT for {}: {}, {}, {}, {}", operatorPortName, s.tupleCount, portElapsedMillis, operatorPortLastEndWindowTimestamps.get(operatorPortName), lastStatsTimestamp);
ps.tuplesPMSMA.add(s.tupleCount, portElapsedMillis);
ps.bufferServerBytesPMSMA.add(s.bufferServerBytes, portElapsedMillis);
ps.queueSizeMA.add(s.queueSize);
operatorPortLastEndWindowTimestamps.put(operatorPortName, s.endWindowTimestamp);
if (maxEndWindowTimestamp < s.endWindowTimestamp) {
maxEndWindowTimestamp = s.endWindowTimestamp;
}
if (s.endWindowTimestamp > maxDequeueTimestamp) {
maxDequeueTimestamp = s.endWindowTimestamp;
}
}
// need to remove dead ports, for unifiers
Iterator<Map.Entry<String, PortStatus>> it = status.inputPortStatusList.entrySet().iterator();
while (it.hasNext()) {
Map.Entry<String, PortStatus> entry = it.next();
if (!currentInputPortSet.contains(entry.getKey())) {
it.remove();
}
}
}
ports = stats.outputPorts;
if (ports != null) {
Set<String> currentOutputPortSet = Sets.newHashSetWithExpectedSize(ports.size());
for (ContainerStats.OperatorStats.PortStats s : ports) {
currentOutputPortSet.add(s.id);
PortStatus ps = status.outputPortStatusList.get(s.id);
if (ps == null) {
ps = status.new PortStatus();
ps.portName = s.id;
status.outputPortStatusList.put(s.id, ps);
}
ps.totalTuples += s.tupleCount;
ps.recordingId = s.recordingId;
tuplesEmitted += s.tupleCount;
Pair<Integer, String> operatorPortName = new Pair<>(oper.getId(), s.id);
Long lastEndWindowTimestamp = operatorPortLastEndWindowTimestamps.get(operatorPortName);
if (lastEndWindowTimestamp == null) {
lastEndWindowTimestamp = lastStatsTimestamp;
}
long portElapsedMillis = Math.max(s.endWindowTimestamp - lastEndWindowTimestamp, 0);
// LOG.debug("=== EMITTED TUPLE COUNT for {}: {}, {}, {}, {}", operatorPortName, s.tupleCount, portElapsedMillis, operatorPortLastEndWindowTimestamps.get(operatorPortName), lastStatsTimestamp);
ps.tuplesPMSMA.add(s.tupleCount, portElapsedMillis);
ps.bufferServerBytesPMSMA.add(s.bufferServerBytes, portElapsedMillis);
operatorPortLastEndWindowTimestamps.put(operatorPortName, s.endWindowTimestamp);
if (maxEndWindowTimestamp < s.endWindowTimestamp) {
maxEndWindowTimestamp = s.endWindowTimestamp;
}
}
if (ports.size() > 0) {
endWindowStats.emitTimestamp = ports.iterator().next().endWindowTimestamp;
}
// need to remove dead ports, for unifiers
Iterator<Map.Entry<String, PortStatus>> it = status.outputPortStatusList.entrySet().iterator();
while (it.hasNext()) {
Map.Entry<String, PortStatus> entry = it.next();
if (!currentOutputPortSet.contains(entry.getKey())) {
it.remove();
}
}
}
// (we don't know the latency for output operators because they don't emit tuples)
if (endWindowStats.emitTimestamp < 0) {
endWindowStats.emitTimestamp = maxDequeueTimestamp;
}
if (status.currentWindowId.get() != stats.windowId) {
status.lastWindowIdChangeTms = currentTimeMillis;
status.currentWindowId.set(stats.windowId);
}
totalCpuTimeUsed += stats.cpuTimeUsed;
statCount++;
if (oper.getOperatorMeta().getValue(OperatorContext.COUNTERS_AGGREGATOR) != null) {
endWindowStats.counters = stats.counters;
}
if (oper.getOperatorMeta().getMetricAggregatorMeta() != null && oper.getOperatorMeta().getMetricAggregatorMeta().getAggregator() != null) {
endWindowStats.metrics = stats.metrics;
}
if (stats.windowId > currentEndWindowStatsWindowId) {
Map<Integer, EndWindowStats> endWindowStatsMap = endWindowStatsOperatorMap.get(stats.windowId);
if (endWindowStatsMap == null) {
endWindowStatsMap = new ConcurrentSkipListMap<>();
Map<Integer, EndWindowStats> endWindowStatsMapPrevious = endWindowStatsOperatorMap.putIfAbsent(stats.windowId, endWindowStatsMap);
if (endWindowStatsMapPrevious != null) {
endWindowStatsMap = endWindowStatsMapPrevious;
}
}
endWindowStatsMap.put(shb.getNodeId(), endWindowStats);
Set<Integer> allCurrentOperators = plan.getAllOperators().keySet();
int numOperators = plan.getAllOperators().size();
if (allCurrentOperators.containsAll(endWindowStatsMap.keySet()) && endWindowStatsMap.size() == numOperators) {
completeEndWindowStatsWindowId = stats.windowId;
}
}
}
status.totalTuplesProcessed.add(tuplesProcessed);
status.totalTuplesEmitted.add(tuplesEmitted);
OperatorMeta logicalOperator = oper.getOperatorMeta();
LogicalOperatorStatus logicalStatus = logicalOperator.getStatus();
if (!oper.isUnifier()) {
logicalStatus.totalTuplesProcessed += tuplesProcessed;
logicalStatus.totalTuplesEmitted += tuplesEmitted;
}
long lastMaxEndWindowTimestamp = operatorLastEndWindowTimestamps.containsKey(oper.getId()) ? operatorLastEndWindowTimestamps.get(oper.getId()) : lastStatsTimestamp;
if (maxEndWindowTimestamp >= lastMaxEndWindowTimestamp) {
double tuplesProcessedPMSMA = 0.0;
double tuplesEmittedPMSMA = 0.0;
if (statCount != 0) {
// LOG.debug("CPU for {}: {} / {} - {}", oper.getId(), totalCpuTimeUsed, maxEndWindowTimestamp, lastMaxEndWindowTimestamp);
status.cpuNanosPMSMA.add(totalCpuTimeUsed, maxEndWindowTimestamp - lastMaxEndWindowTimestamp);
}
for (PortStatus ps : status.inputPortStatusList.values()) {
tuplesProcessedPMSMA += ps.tuplesPMSMA.getAvg();
}
for (PortStatus ps : status.outputPortStatusList.values()) {
tuplesEmittedPMSMA += ps.tuplesPMSMA.getAvg();
}
status.tuplesProcessedPSMA.set(Math.round(tuplesProcessedPMSMA * 1000));
status.tuplesEmittedPSMA.set(Math.round(tuplesEmittedPMSMA * 1000));
} else {
// LOG.warn("This timestamp for {} is lower than the previous!! {} < {}", oper.getId(),
// maxEndWindowTimestamp, lastMaxEndWindowTimestamp);
}
operatorLastEndWindowTimestamps.put(oper.getId(), maxEndWindowTimestamp);
status.listenerStats.add(statsList);
this.reportStats.put(oper, oper);
status.statsRevs.commit();
}
if (lastStatsTimestamp < maxEndWindowTimestamp) {
lastStatsTimestamp = maxEndWindowTimestamp;
}
}
sca.lastHeartbeatMillis = currentTimeMillis;
for (PTOperator oper : sca.container.getOperators()) {
if (!reportedOperators.contains(oper.getId())) {
processOperatorDeployStatus(oper, null, sca);
}
}
ContainerHeartbeatResponse rsp = getHeartbeatResponse(sca);
if (heartbeat.getContainerStats().operators.isEmpty() && isApplicationIdle()) {
LOG.info("requesting idle shutdown for container {}", heartbeat.getContainerId());
rsp.shutdown = ShutdownType.ABORT;
} else {
if (sca.isShutdownRequested()) {
LOG.info("requesting shutdown for container {}", heartbeat.getContainerId());
rsp.shutdown = sca.shutdownRequest;
}
}
List<StramToNodeRequest> requests = rsp.nodeRequests != null ? rsp.nodeRequests : new ArrayList<StramToNodeRequest>();
ConcurrentLinkedQueue<StramToNodeRequest> operatorRequests = sca.getOperatorRequests();
while (true) {
StramToNodeRequest r = operatorRequests.poll();
if (r == null) {
break;
}
requests.add(r);
}
rsp.nodeRequests = requests;
rsp.committedWindowId = committedWindowId;
rsp.stackTraceRequired = sca.stackTraceRequested;
sca.stackTraceRequested = false;
apexPluginDispatcher.dispatch(new DAGExecutionEvent.HeartbeatExecutionEvent(heartbeat));
return rsp;
}
Aggregations