@Override
protected void map(Object key, Object value, Context context)
throws IOException, InterruptedException {
try {
final InputRow inputRow;
try {
inputRow = parseInputRow(value, parser);
} catch (Exception e) {
if (config.isIgnoreInvalidRows()) {
log.debug(e, "Ignoring invalid row [%s] due to parsing error", value.toString());
context
.getCounter(HadoopDruidIndexerConfig.IndexJobCounters.INVALID_ROW_COUNTER)
.increment(1);
return; // we're ignoring this invalid row
} else {
throw e;
}
}
if (!granularitySpec.bucketIntervals().isPresent()
|| granularitySpec
.bucketInterval(new DateTime(inputRow.getTimestampFromEpoch()))
.isPresent()) {
innerMap(inputRow, value, context);
}
} catch (RuntimeException e) {
throw new RE(e, "Failure on row[%s]", value);
}
}
@Override
protected void innerMap(InputRow inputRow, Object value, Context context)
throws IOException, InterruptedException {
final Map<String, Iterable<String>> dims = Maps.newHashMap();
for (final String dim : inputRow.getDimensions()) {
dims.put(dim, inputRow.getDimension(dim));
}
helper.emitDimValueCounts(context, new DateTime(inputRow.getTimestampFromEpoch()), dims);
}
@Override
protected void innerMap(InputRow inputRow, Object value, Context context)
throws IOException, InterruptedException {
final List<Object> groupKey =
Rows.toGroupKey(rollupGranularity.truncate(inputRow.getTimestampFromEpoch()), inputRow);
context.write(
new BytesWritable(HadoopDruidIndexerConfig.JSON_MAPPER.writeValueAsBytes(groupKey)),
NullWritable.get());
}
@Test(timeout = 60000)
public void testPersistFails() throws Exception {
final MutableBoolean committed = new MutableBoolean(false);
plumber
.getSinks()
.put(
0L,
new Sink(
new Interval(0, TimeUnit.HOURS.toMillis(1)),
schema,
tuningConfig,
new DateTime("2014-12-01T12:34:56.789").toString()));
plumber.startJob();
final InputRow row = EasyMock.createNiceMock(InputRow.class);
EasyMock.expect(row.getTimestampFromEpoch()).andReturn(0L);
EasyMock.expect(row.getDimensions()).andReturn(new ArrayList<String>());
EasyMock.replay(row);
plumber.add(row, Committers.supplierOf(Committers.nil()));
plumber.persist(
Committers.supplierFromRunnable(
new Runnable() {
@Override
public void run() {
committed.setValue(true);
throw new RuntimeException();
}
})
.get());
while (!committed.booleanValue()) {
Thread.sleep(100);
}
// Exception may need time to propagate
while (metrics.failedPersists() < 1) {
Thread.sleep(100);
}
Assert.assertEquals(1, metrics.failedPersists());
}
private void testPersist(final Object commitMetadata) throws Exception {
final MutableBoolean committed = new MutableBoolean(false);
plumber
.getSinks()
.put(
0L,
new Sink(
new Interval(0, TimeUnit.HOURS.toMillis(1)),
schema,
tuningConfig,
new DateTime("2014-12-01T12:34:56.789").toString()));
Assert.assertNull(plumber.startJob());
final InputRow row = EasyMock.createNiceMock(InputRow.class);
EasyMock.expect(row.getTimestampFromEpoch()).andReturn(0L);
EasyMock.expect(row.getDimensions()).andReturn(new ArrayList<String>());
EasyMock.replay(row);
final Committer committer =
new Committer() {
@Override
public Object getMetadata() {
return commitMetadata;
}
@Override
public void run() {
committed.setValue(true);
}
};
plumber.add(row, Suppliers.ofInstance(committer));
plumber.persist(committer);
while (!committed.booleanValue()) {
Thread.sleep(100);
}
plumber.getSinks().clear();
plumber.finishJob();
}
@Override
public void run() {
plumber = initPlumber();
final Period intermediatePersistPeriod = config.getIntermediatePersistPeriod();
try {
plumber.startJob();
// Delay firehose connection to avoid claiming input resources while the plumber is starting
// up.
firehose = initFirehose();
long nextFlush = new DateTime().plus(intermediatePersistPeriod).getMillis();
while (firehose.hasMore()) {
InputRow inputRow = null;
try {
try {
inputRow = firehose.nextRow();
} catch (Exception e) {
log.debug(e, "thrown away line due to exception, considering unparseable");
metrics.incrementUnparseable();
continue;
}
boolean lateEvent = false;
boolean indexLimitExceeded = false;
try {
lateEvent = plumber.add(inputRow) == -1;
} catch (IndexSizeExceededException e) {
log.info("Index limit exceeded: %s", e.getMessage());
indexLimitExceeded = true;
}
if (indexLimitExceeded || lateEvent) {
metrics.incrementThrownAway();
log.debug("Throwing away event[%s]", inputRow);
if (indexLimitExceeded || System.currentTimeMillis() > nextFlush) {
plumber.persist(firehose.commit());
nextFlush = new DateTime().plus(intermediatePersistPeriod).getMillis();
}
continue;
}
final Sink sink = plumber.getSink(inputRow.getTimestampFromEpoch());
if ((sink != null && !sink.canAppendRow()) || System.currentTimeMillis() > nextFlush) {
plumber.persist(firehose.commit());
nextFlush = new DateTime().plus(intermediatePersistPeriod).getMillis();
}
metrics.incrementProcessed();
} catch (ParseException e) {
if (inputRow != null) {
log.error(e, "unparseable line: %s", inputRow);
}
metrics.incrementUnparseable();
}
}
} catch (RuntimeException e) {
log.makeAlert(
e,
"RuntimeException aborted realtime processing[%s]",
fireDepartment.getDataSchema().getDataSource())
.emit();
normalExit = false;
throw e;
} catch (Error e) {
log.makeAlert(
e,
"Exception aborted realtime processing[%s]",
fireDepartment.getDataSchema().getDataSource())
.emit();
normalExit = false;
throw e;
} finally {
CloseQuietly.close(firehose);
if (normalExit) {
plumber.finishJob();
plumber = null;
firehose = null;
}
}
}
@Override
public TaskStatus run(final TaskToolbox toolbox) throws Exception {
if (this.plumber != null) {
throw new IllegalStateException("WTF?!? run with non-null plumber??!");
}
// Shed any locks we might have (e.g. if we were uncleanly killed and restarted) since we'll
// reacquire
// them if we actually need them
for (final TaskLock taskLock : getTaskLocks(toolbox)) {
toolbox.getTaskActionClient().submit(new LockReleaseAction(taskLock.getInterval()));
}
boolean normalExit = true;
// Set up firehose
final Period intermediatePersistPeriod = fireDepartmentConfig.getIntermediatePersistPeriod();
final Firehose firehose = firehoseFactory.connect();
// It would be nice to get the PlumberSchool in the constructor. Although that will need
// jackson injectables for
// stuff like the ServerView, which seems kind of odd? Perhaps revisit this when Guice has been
// introduced.
final RealtimePlumberSchool realtimePlumberSchool =
new RealtimePlumberSchool(
windowPeriod, new File(toolbox.getTaskWorkDir(), "persist"), segmentGranularity);
realtimePlumberSchool.setDefaultMaxPendingPersists(maxPendingPersists);
final SegmentPublisher segmentPublisher = new TaskActionSegmentPublisher(this, toolbox);
// NOTE: We talk to the coordinator in various places in the plumber and we could be more robust
// to issues
// with the coordinator. Right now, we'll block/throw in whatever thread triggered the
// coordinator behavior,
// which will typically be either the main data processing loop or the persist thread.
// Wrap default DataSegmentAnnouncer such that we unlock intervals as we unannounce segments
final DataSegmentAnnouncer lockingSegmentAnnouncer =
new DataSegmentAnnouncer() {
@Override
public void announceSegment(final DataSegment segment) throws IOException {
// Side effect: Calling announceSegment causes a lock to be acquired
toolbox.getTaskActionClient().submit(new LockAcquireAction(segment.getInterval()));
toolbox.getSegmentAnnouncer().announceSegment(segment);
}
@Override
public void unannounceSegment(final DataSegment segment) throws IOException {
try {
toolbox.getSegmentAnnouncer().unannounceSegment(segment);
} finally {
toolbox.getTaskActionClient().submit(new LockReleaseAction(segment.getInterval()));
}
}
@Override
public void announceSegments(Iterable<DataSegment> segments) throws IOException {
// Side effect: Calling announceSegments causes locks to be acquired
for (DataSegment segment : segments) {
toolbox.getTaskActionClient().submit(new LockAcquireAction(segment.getInterval()));
}
toolbox.getSegmentAnnouncer().announceSegments(segments);
}
@Override
public void unannounceSegments(Iterable<DataSegment> segments) throws IOException {
try {
toolbox.getSegmentAnnouncer().unannounceSegments(segments);
} finally {
for (DataSegment segment : segments) {
toolbox.getTaskActionClient().submit(new LockReleaseAction(segment.getInterval()));
}
}
}
};
// NOTE: getVersion will block if there is lock contention, which will block plumber.getSink
// NOTE: (and thus the firehose)
// Shouldn't usually happen, since we don't expect people to submit tasks that intersect with
// the
// realtime window, but if they do it can be problematic. If we decide to care, we can use more
// threads in
// the plumber such that waiting for the coordinator doesn't block data processing.
final VersioningPolicy versioningPolicy =
new VersioningPolicy() {
@Override
public String getVersion(final Interval interval) {
try {
// Side effect: Calling getVersion causes a lock to be acquired
final TaskLock myLock =
toolbox.getTaskActionClient().submit(new LockAcquireAction(interval));
return myLock.getVersion();
} catch (IOException e) {
throw Throwables.propagate(e);
}
}
};
// NOTE: This pusher selects path based purely on global configuration and the DataSegment,
// which means
// NOTE: that redundant realtime tasks will upload to the same location. This can cause
// index.zip and
// NOTE: descriptor.json to mismatch, or it can cause historical nodes to load different
// instances of the
// NOTE: "same" segment.
realtimePlumberSchool.setDataSegmentPusher(toolbox.getSegmentPusher());
realtimePlumberSchool.setConglomerate(toolbox.getQueryRunnerFactoryConglomerate());
realtimePlumberSchool.setQueryExecutorService(toolbox.getQueryExecutorService());
realtimePlumberSchool.setVersioningPolicy(versioningPolicy);
realtimePlumberSchool.setSegmentAnnouncer(lockingSegmentAnnouncer);
realtimePlumberSchool.setSegmentPublisher(segmentPublisher);
realtimePlumberSchool.setServerView(toolbox.getNewSegmentServerView());
realtimePlumberSchool.setEmitter(toolbox.getEmitter());
if (this.rejectionPolicyFactory != null) {
realtimePlumberSchool.setRejectionPolicyFactory(rejectionPolicyFactory);
}
final FireDepartment fireDepartment =
new FireDepartment(schema, fireDepartmentConfig, null, null);
final RealtimeMetricsMonitor metricsMonitor =
new RealtimeMetricsMonitor(ImmutableList.of(fireDepartment));
this.queryRunnerFactoryConglomerate = toolbox.getQueryRunnerFactoryConglomerate();
this.plumber = realtimePlumberSchool.findPlumber(schema, fireDepartment.getMetrics());
try {
plumber.startJob();
// Set up metrics emission
toolbox.getMonitorScheduler().addMonitor(metricsMonitor);
// Time to read data!
long nextFlush = new DateTime().plus(intermediatePersistPeriod).getMillis();
while (firehose.hasMore()) {
final InputRow inputRow;
try {
inputRow = firehose.nextRow();
if (inputRow == null) {
continue;
}
final Sink sink = plumber.getSink(inputRow.getTimestampFromEpoch());
if (sink == null) {
fireDepartment.getMetrics().incrementThrownAway();
log.debug("Throwing away event[%s]", inputRow);
if (System.currentTimeMillis() > nextFlush) {
plumber.persist(firehose.commit());
nextFlush = new DateTime().plus(intermediatePersistPeriod).getMillis();
}
continue;
}
if (sink.isEmpty()) {
log.info("Task %s: New sink: %s", getId(), sink);
}
int currCount = sink.add(inputRow);
fireDepartment.getMetrics().incrementProcessed();
if (currCount >= fireDepartmentConfig.getMaxRowsInMemory()
|| System.currentTimeMillis() > nextFlush) {
plumber.persist(firehose.commit());
nextFlush = new DateTime().plus(intermediatePersistPeriod).getMillis();
}
} catch (FormattedException e) {
log.warn(e, "unparseable line");
fireDepartment.getMetrics().incrementUnparseable();
}
}
} catch (Throwable e) {
normalExit = false;
log.makeAlert(e, "Exception aborted realtime processing[%s]", schema.getDataSource()).emit();
throw e;
} finally {
if (normalExit) {
try {
plumber.persist(firehose.commit());
plumber.finishJob();
} catch (Exception e) {
log.makeAlert(e, "Failed to finish realtime task").emit();
} finally {
Closeables.closeQuietly(firehose);
toolbox.getMonitorScheduler().removeMonitor(metricsMonitor);
}
}
}
return TaskStatus.success(getId());
}