Examples with Sink org.apache.druid.segment.realtime.plumber.Sink used on opensource projects

Example 1 with Sink

use of org.apache.druid.segment.realtime.plumber.Sink in project druid by druid-io.

the class BatchAppenderator method add.

@Override
public AppenderatorAddResult add(final SegmentIdWithShardSpec identifier, final InputRow row, @Nullable final Supplier<Committer> committerSupplier, final boolean allowIncrementalPersists) throws IndexSizeExceededException, SegmentNotWritableException {
    throwPersistErrorIfExists();
    Preconditions.checkArgument(committerSupplier == null, "Batch appenderator does not need a committer!");
    Preconditions.checkArgument(allowIncrementalPersists, "Batch appenderator should always allow incremental persists!");
    if (!identifier.getDataSource().equals(schema.getDataSource())) {
        throw new IAE("Expected dataSource[%s] but was asked to insert row for dataSource[%s]?!", schema.getDataSource(), identifier.getDataSource());
    }
    final Sink sink = getOrCreateSink(identifier);
    metrics.reportMessageMaxTimestamp(row.getTimestampFromEpoch());
    final int sinkRowsInMemoryBeforeAdd = sink.getNumRowsInMemory();
    final int sinkRowsInMemoryAfterAdd;
    final long bytesInMemoryBeforeAdd = sink.getBytesInMemory();
    final long bytesInMemoryAfterAdd;
    final IncrementalIndexAddResult addResult;
    try {
        // allow incrememtal persis is always true for batch
        addResult = sink.add(row, false);
        sinkRowsInMemoryAfterAdd = addResult.getRowCount();
        bytesInMemoryAfterAdd = addResult.getBytesInMemory();
    } catch (IndexSizeExceededException e) {
        // Uh oh, we can't do anything about this! We can't persist (commit metadata would be out of sync) and we
        // can't add the row (it just failed). This should never actually happen, though, because we check
        // sink.canAddRow after returning from add.
        log.error(e, "Sink for segment[%s] was unexpectedly full!", identifier);
        throw e;
    }
    if (sinkRowsInMemoryAfterAdd < 0) {
        throw new SegmentNotWritableException("Attempt to add row to swapped-out sink for segment[%s].", identifier);
    }
    if (addResult.isRowAdded()) {
        rowIngestionMeters.incrementProcessed();
    } else if (addResult.hasParseException()) {
        parseExceptionHandler.handle(addResult.getParseException());
    }
    final int numAddedRows = sinkRowsInMemoryAfterAdd - sinkRowsInMemoryBeforeAdd;
    rowsCurrentlyInMemory += numAddedRows;
    bytesCurrentlyInMemory += (bytesInMemoryAfterAdd - bytesInMemoryBeforeAdd);
    totalRows += numAddedRows;
    sinksMetadata.computeIfAbsent(identifier, unused -> new SinkMetadata()).addRows(numAddedRows);
    boolean persist = false;
    List<String> persistReasons = new ArrayList<>();
    if (!sink.canAppendRow()) {
        persist = true;
        persistReasons.add("No more rows can be appended to sink");
    }
    if (rowsCurrentlyInMemory >= tuningConfig.getMaxRowsInMemory()) {
        persist = true;
        persistReasons.add(StringUtils.format("rowsCurrentlyInMemory[%d] is greater than maxRowsInMemory[%d]", rowsCurrentlyInMemory, tuningConfig.getMaxRowsInMemory()));
    }
    if (bytesCurrentlyInMemory >= maxBytesTuningConfig) {
        persist = true;
        persistReasons.add(StringUtils.format("bytesCurrentlyInMemory[%d] is greater than maxBytesInMemory[%d]", bytesCurrentlyInMemory, maxBytesTuningConfig));
    }
    if (persist) {
        // persistAll clears rowsCurrentlyInMemory, no need to update it.
        log.info("Incremental persist to disk because %s.", String.join(",", persistReasons));
        long bytesToBePersisted = 0L;
        for (Map.Entry<SegmentIdWithShardSpec, Sink> entry : sinks.entrySet()) {
            final Sink sinkEntry = entry.getValue();
            if (sinkEntry != null) {
                bytesToBePersisted += sinkEntry.getBytesInMemory();
                if (sinkEntry.swappable()) {
                    // Code for batch no longer memory maps hydrants, but they still take memory...
                    int memoryStillInUse = calculateMemoryUsedByHydrant();
                    bytesCurrentlyInMemory += memoryStillInUse;
                }
            }
        }
        if (!skipBytesInMemoryOverheadCheck && bytesCurrentlyInMemory - bytesToBePersisted > maxBytesTuningConfig) {
            // We are still over maxBytesTuningConfig even after persisting.
            // This means that we ran out of all available memory to ingest (due to overheads created as part of ingestion)
            final String alertMessage = StringUtils.format("Task has exceeded safe estimated heap usage limits, failing " + "(numSinks: [%d] numHydrantsAcrossAllSinks: [%d] totalRows: [%d])" + "(bytesCurrentlyInMemory: [%d] - bytesToBePersisted: [%d] > maxBytesTuningConfig: [%d])", sinks.size(), sinks.values().stream().mapToInt(Iterables::size).sum(), getTotalRowCount(), bytesCurrentlyInMemory, bytesToBePersisted, maxBytesTuningConfig);
            final String errorMessage = StringUtils.format("%s.\nThis can occur when the overhead from too many intermediary segment persists becomes to " + "great to have enough space to process additional input rows. This check, along with metering the overhead " + "of these objects to factor into the 'maxBytesInMemory' computation, can be disabled by setting " + "'skipBytesInMemoryOverheadCheck' to 'true' (note that doing so might allow the task to naturally encounter " + "a 'java.lang.OutOfMemoryError'). Alternatively, 'maxBytesInMemory' can be increased which will cause an " + "increase in heap footprint, but will allow for more intermediary segment persists to occur before " + "reaching this condition.", alertMessage);
            log.makeAlert(alertMessage).addData("dataSource", schema.getDataSource()).emit();
            throw new RuntimeException(errorMessage);
        }
        Futures.addCallback(persistAll(null), new FutureCallback<Object>() {

            @Override
            public void onSuccess(@Nullable Object result) {
            // do nothing
            }

            @Override
            public void onFailure(Throwable t) {
                persistError = t;
            }
        });
    }
    return new AppenderatorAddResult(identifier, sinksMetadata.get(identifier).numRowsInSegment, false);
}

Example 2 with Sink

use of org.apache.druid.segment.realtime.plumber.Sink in project druid by druid-io.

the class BatchAppenderator method persistAll.

@Override
public ListenableFuture<Object> persistAll(@Nullable final Committer committer) {
    throwPersistErrorIfExists();
    if (committer != null) {
        throw new ISE("committer must be null for BatchAppenderator");
    }
    // Get ready to persist all sinks:
    final Map<SegmentIdWithShardSpec, Sink> sinksToPersist = swapSinks();
    final Stopwatch runExecStopwatch = Stopwatch.createStarted();
    ListenableFuture<Object> future = persistExecutor.submit(() -> {
        log.info("Spawning intermediate persist");
        // figure out hydrants (indices) to persist:
        final List<Pair<FireHydrant, SegmentIdWithShardSpec>> indexesToPersist = new ArrayList<>();
        int numPersistedRows = 0;
        long bytesPersisted = 0;
        int totalHydrantsCount = 0;
        final long totalSinks = sinksToPersist.size();
        for (Map.Entry<SegmentIdWithShardSpec, Sink> entry : sinksToPersist.entrySet()) {
            final SegmentIdWithShardSpec identifier = entry.getKey();
            final Sink sink = entry.getValue();
            if (sink == null) {
                throw new ISE("No sink for identifier: %s", identifier);
            }
            final List<FireHydrant> hydrants = Lists.newArrayList(sink);
            // Since everytime we persist we also get rid of the in-memory references to sink & hydrants
            // the invariant of exactly one, always swappable, sink with exactly one unpersisted hydrant must hold
            int totalHydrantsForSink = hydrants.size();
            if (totalHydrantsForSink != 1) {
                throw new ISE("There should be only one hydrant for identifier[%s] but there are[%s]", identifier, totalHydrantsForSink);
            }
            totalHydrantsCount++;
            numPersistedRows += sink.getNumRowsInMemory();
            bytesPersisted += sink.getBytesInMemory();
            if (!sink.swappable()) {
                throw new ISE("Sink is not swappable![%s]", identifier);
            }
            indexesToPersist.add(Pair.of(sink.swap(), identifier));
        }
        if (indexesToPersist.isEmpty()) {
            log.info("No indexes will be persisted");
        }
        final Stopwatch persistStopwatch = Stopwatch.createStarted();
        try {
            for (Pair<FireHydrant, SegmentIdWithShardSpec> pair : indexesToPersist) {
                metrics.incrementRowOutputCount(persistHydrant(pair.lhs, pair.rhs));
            }
            log.info("Persisted in-memory data for segments: %s", indexesToPersist.stream().filter(itp -> itp.rhs != null).map(itp -> itp.rhs.asSegmentId().toString()).distinct().collect(Collectors.joining(", ")));
            log.info("Persisted stats: processed rows: [%d], persisted rows[%d], persisted sinks: [%d], persisted fireHydrants (across sinks): [%d]", rowIngestionMeters.getProcessed(), numPersistedRows, totalSinks, totalHydrantsCount);
        // note that we do not need to reset sinks metadata since we did it at the start...
        } catch (Exception e) {
            metrics.incrementFailedPersists();
            throw e;
        } finally {
            metrics.incrementNumPersists();
            long persistMillis = persistStopwatch.elapsed(TimeUnit.MILLISECONDS);
            metrics.incrementPersistTimeMillis(persistMillis);
            persistStopwatch.stop();
            // make sure no push can start while persisting:
            log.info("Persisted rows[%,d] and bytes[%,d] and removed all sinks & hydrants from memory in[%d] millis", numPersistedRows, bytesPersisted, persistMillis);
            log.info("Persist is done.");
        }
        return null;
    });
    final long startDelay = runExecStopwatch.elapsed(TimeUnit.MILLISECONDS);
    metrics.incrementPersistBackPressureMillis(startDelay);
    if (startDelay > PERSIST_WARN_DELAY) {
        log.warn("Ingestion was throttled for [%,d] millis because persists were pending.", startDelay);
    }
    runExecStopwatch.stop();
    return future;
}

Example 3 with Sink

use of org.apache.druid.segment.realtime.plumber.Sink in project druid by druid-io.

the class BatchAppenderator method getOrCreateSink.

private Sink getOrCreateSink(final SegmentIdWithShardSpec identifier) {
    Sink retVal = sinks.get(identifier);
    if (retVal == null) {
        retVal = new Sink(identifier.getInterval(), schema, identifier.getShardSpec(), identifier.getVersion(), tuningConfig.getAppendableIndexSpec(), tuningConfig.getMaxRowsInMemory(), maxBytesTuningConfig, useMaxMemoryEstimates, null);
        bytesCurrentlyInMemory += calculateSinkMemoryInUsed();
        sinks.put(identifier, retVal);
        metrics.setSinkCount(sinks.size());
    }
    return retVal;
}

Example 4 with Sink

use of org.apache.druid.segment.realtime.plumber.Sink in project druid by druid-io.

the class StreamAppenderator method push.

@Override
public ListenableFuture<SegmentsAndCommitMetadata> push(final Collection<SegmentIdWithShardSpec> identifiers, @Nullable final Committer committer, final boolean useUniquePath) {
    final Map<SegmentIdWithShardSpec, Sink> theSinks = new HashMap<>();
    AtomicLong pushedHydrantsCount = new AtomicLong();
    for (final SegmentIdWithShardSpec identifier : identifiers) {
        final Sink sink = sinks.get(identifier);
        if (sink == null) {
            throw new ISE("No sink for identifier: %s", identifier);
        }
        theSinks.put(identifier, sink);
        if (sink.finishWriting()) {
            totalRows.addAndGet(-sink.getNumRows());
        }
        // count hydrants for stats:
        pushedHydrantsCount.addAndGet(Iterables.size(sink));
    }
    return Futures.transform(// segments.
    persistAll(committer), (Function<Object, SegmentsAndCommitMetadata>) commitMetadata -> {
        final List<DataSegment> dataSegments = new ArrayList<>();
        log.info("Preparing to push (stats): processed rows: [%d], sinks: [%d], fireHydrants (across sinks): [%d]", rowIngestionMeters.getProcessed(), theSinks.size(), pushedHydrantsCount.get());
        log.debug("Building and pushing segments: %s", theSinks.keySet().stream().map(SegmentIdWithShardSpec::toString).collect(Collectors.joining(", ")));
        for (Map.Entry<SegmentIdWithShardSpec, Sink> entry : theSinks.entrySet()) {
            if (droppingSinks.contains(entry.getKey())) {
                log.warn("Skipping push of currently-dropping sink[%s]", entry.getKey());
                continue;
            }
            final DataSegment dataSegment = mergeAndPush(entry.getKey(), entry.getValue(), useUniquePath);
            if (dataSegment != null) {
                dataSegments.add(dataSegment);
            } else {
                log.warn("mergeAndPush[%s] returned null, skipping.", entry.getKey());
            }
        }
        log.info("Push complete...");
        return new SegmentsAndCommitMetadata(dataSegments, commitMetadata);
    }, pushExecutor);
}

Example 5 with Sink

use of org.apache.druid.segment.realtime.plumber.Sink in project druid by druid-io.

the class StreamAppenderator method close.

@Override
public void close() {
    if (!closed.compareAndSet(false, true)) {
        log.debug("Appenderator already closed, skipping close() call.");
        return;
    }
    log.debug("Shutting down...");
    final List<ListenableFuture<?>> futures = new ArrayList<>();
    for (Map.Entry<SegmentIdWithShardSpec, Sink> entry : sinks.entrySet()) {
        futures.add(abandonSegment(entry.getKey(), entry.getValue(), false));
    }
    try {
        Futures.allAsList(futures).get();
    } catch (InterruptedException e) {
        Thread.currentThread().interrupt();
        log.warn(e, "Interrupted during close()");
    } catch (ExecutionException e) {
        log.warn(e, "Unable to abandon existing segments during close()");
    }
    try {
        shutdownExecutors();
        Preconditions.checkState(persistExecutor == null || persistExecutor.awaitTermination(365, TimeUnit.DAYS), "persistExecutor not terminated");
        Preconditions.checkState(pushExecutor == null || pushExecutor.awaitTermination(365, TimeUnit.DAYS), "pushExecutor not terminated");
        Preconditions.checkState(intermediateTempExecutor == null || intermediateTempExecutor.awaitTermination(365, TimeUnit.DAYS), "intermediateTempExecutor not terminated");
        persistExecutor = null;
        pushExecutor = null;
        intermediateTempExecutor = null;
    } catch (InterruptedException e) {
        Thread.currentThread().interrupt();
        throw new ISE("Failed to shutdown executors during close()");
    }
    // Only unlock if executors actually shut down.
    unlockBasePersistDirectory();
}