Examples with Segments org.apache.druid.indexing.overlord.Segments used on opensource projects

Example 6 with Segments

use of org.apache.druid.indexing.overlord.Segments in project druid by druid-io.

the class AbstractBatchIndexTask method compactionStateAnnotateFunction.

public static Function<Set<DataSegment>, Set<DataSegment>> compactionStateAnnotateFunction(boolean storeCompactionState, TaskToolbox toolbox, IngestionSpec ingestionSpec) {
    if (storeCompactionState) {
        TuningConfig tuningConfig = ingestionSpec.getTuningConfig();
        GranularitySpec granularitySpec = ingestionSpec.getDataSchema().getGranularitySpec();
        // We do not need to store dimensionExclusions and spatialDimensions since auto compaction does not support them
        DimensionsSpec dimensionsSpec = ingestionSpec.getDataSchema().getDimensionsSpec() == null ? null : new DimensionsSpec(ingestionSpec.getDataSchema().getDimensionsSpec().getDimensions());
        // We only need to store filter since that is the only field auto compaction support
        Map<String, Object> transformSpec = ingestionSpec.getDataSchema().getTransformSpec() == null || TransformSpec.NONE.equals(ingestionSpec.getDataSchema().getTransformSpec()) ? null : new ClientCompactionTaskTransformSpec(ingestionSpec.getDataSchema().getTransformSpec().getFilter()).asMap(toolbox.getJsonMapper());
        List<Object> metricsSpec = ingestionSpec.getDataSchema().getAggregators() == null ? null : toolbox.getJsonMapper().convertValue(ingestionSpec.getDataSchema().getAggregators(), new TypeReference<List<Object>>() {
        });
        final CompactionState compactionState = new CompactionState(tuningConfig.getPartitionsSpec(), dimensionsSpec, metricsSpec, transformSpec, tuningConfig.getIndexSpec().asMap(toolbox.getJsonMapper()), granularitySpec.asMap(toolbox.getJsonMapper()));
        return segments -> segments.stream().map(s -> s.withLastCompactionState(compactionState)).collect(Collectors.toSet());
    } else {
        return Function.identity();
    }
}

Example 7 with Segments

use of org.apache.druid.indexing.overlord.Segments in project druid by druid-io.

the class AbstractBatchIndexTask method findInputSegments.

/**
 * If the given firehoseFactory is {@link IngestSegmentFirehoseFactory}, then it finds the segments to lock
 * from the firehoseFactory. This is because those segments will be read by this task no matter what segments would be
 * filtered by intervalsToRead, so they need to be locked.
 * <p>
 * However, firehoseFactory is not IngestSegmentFirehoseFactory, it means this task will overwrite some segments
 * with data read from some input source outside of Druid. As a result, only the segments falling in intervalsToRead
 * should be locked.
 * <p>
 * The order of segments within the returned list is unspecified, but each segment is guaranteed to appear in the list
 * only once.
 */
protected static List<DataSegment> findInputSegments(String dataSource, TaskActionClient actionClient, List<Interval> intervalsToRead, FirehoseFactory firehoseFactory) throws IOException {
    if (firehoseFactory instanceof IngestSegmentFirehoseFactory) {
        // intervalsToRead is ignored here.
        final List<WindowedSegmentId> inputSegments = ((IngestSegmentFirehoseFactory) firehoseFactory).getSegments();
        if (inputSegments == null) {
            final Interval inputInterval = Preconditions.checkNotNull(((IngestSegmentFirehoseFactory) firehoseFactory).getInterval(), "input interval");
            return ImmutableList.copyOf(actionClient.submit(new RetrieveUsedSegmentsAction(dataSource, inputInterval, null, Segments.ONLY_VISIBLE)));
        } else {
            final List<String> inputSegmentIds = inputSegments.stream().map(WindowedSegmentId::getSegmentId).collect(Collectors.toList());
            final Collection<DataSegment> dataSegmentsInIntervals = actionClient.submit(new RetrieveUsedSegmentsAction(dataSource, null, inputSegments.stream().flatMap(windowedSegmentId -> windowedSegmentId.getIntervals().stream()).collect(Collectors.toSet()), Segments.ONLY_VISIBLE));
            return dataSegmentsInIntervals.stream().filter(segment -> inputSegmentIds.contains(segment.getId().toString())).collect(Collectors.toList());
        }
    } else {
        return ImmutableList.copyOf(actionClient.submit(new RetrieveUsedSegmentsAction(dataSource, null, intervalsToRead, Segments.ONLY_VISIBLE)));
    }
}

Example 8 with Segments

use of org.apache.druid.indexing.overlord.Segments in project druid by druid-io.

the class IndexerSQLMetadataStorageCoordinator method createNewSegment.

/**
 * This function creates a new segment for the given datasource/interval/etc. A critical
 * aspect of the creation is to make sure that the new version & new partition number will make
 * sense given the existing segments & pending segments also very important is to avoid
 * clashes with existing pending & used/unused segments.
 * @param handle Database handle
 * @param dataSource datasource for the new segment
 * @param interval interval for the new segment
 * @param partialShardSpec Shard spec info minus segment id stuff
 * @param existingVersion Version of segments in interval, used to compute the version of the very first segment in
 *                        interval
 * @return
 * @throws IOException
 */
@Nullable
private SegmentIdWithShardSpec createNewSegment(final Handle handle, final String dataSource, final Interval interval, final PartialShardSpec partialShardSpec, final String existingVersion) throws IOException {
    // Get the time chunk and associated data segments for the given interval, if any
    final List<TimelineObjectHolder<String, DataSegment>> existingChunks = getTimelineForIntervalsWithHandle(handle, dataSource, ImmutableList.of(interval)).lookup(interval);
    if (existingChunks.size() > 1) {
        // Not possible to expand more than one chunk with a single segment.
        log.warn("Cannot allocate new segment for dataSource[%s], interval[%s]: already have [%,d] chunks.", dataSource, interval, existingChunks.size());
        return null;
    } else {
        // max partitionId of the shardSpecs which share the same partition space.
        SegmentIdWithShardSpec maxId = null;
        if (!existingChunks.isEmpty()) {
            TimelineObjectHolder<String, DataSegment> existingHolder = Iterables.getOnlyElement(existingChunks);
            // noinspection ConstantConditions
            for (DataSegment segment : FluentIterable.from(existingHolder.getObject()).transform(PartitionChunk::getObject).filter(segment -> segment.getShardSpec().sharePartitionSpace(partialShardSpec))) {
                // Note that this will compute the max id of existing, visible, data segments in the time chunk:
                if (maxId == null || maxId.getShardSpec().getPartitionNum() < segment.getShardSpec().getPartitionNum()) {
                    maxId = SegmentIdWithShardSpec.fromDataSegment(segment);
                }
            }
        }
        // Get the version of the existing chunk, we might need it in some of the cases below
        // to compute the new identifier's version
        @Nullable final String versionOfExistingChunk;
        if (!existingChunks.isEmpty()) {
            // remember only one chunk possible for given interval so get the first & only one
            versionOfExistingChunk = existingChunks.get(0).getVersion();
        } else {
            versionOfExistingChunk = null;
        }
        // next, we need to enrich the maxId computed before with the information of the pending segments
        // it is possible that a pending segment has a higher id in which case we need that, it will work,
        // and it will avoid clashes when inserting the new pending segment later in the caller of this method
        final Set<SegmentIdWithShardSpec> pendings = getPendingSegmentsForIntervalWithHandle(handle, dataSource, interval);
        // Make sure we add the maxId we obtained from the segments table:
        if (maxId != null) {
            pendings.add(maxId);
        }
        // Now compute the maxId with all the information: pendings + segments:
        // The versionOfExistingChunks filter is ensure that we pick the max id with the version of the existing chunk
        // in the case that there may be a pending segment with a higher version but no corresponding used segments
        // which may generate a clash with an existing segment once the new id is generated
        maxId = pendings.stream().filter(id -> id.getShardSpec().sharePartitionSpace(partialShardSpec)).filter(id -> versionOfExistingChunk == null ? true : id.getVersion().equals(versionOfExistingChunk)).max((id1, id2) -> {
            final int versionCompare = id1.getVersion().compareTo(id2.getVersion());
            if (versionCompare != 0) {
                return versionCompare;
            } else {
                return Integer.compare(id1.getShardSpec().getPartitionNum(), id2.getShardSpec().getPartitionNum());
            }
        }).orElse(null);
        // The following code attempts to compute the new version, if this
        // new version is not null at the end of next block then it will be
        // used as the new version in the case for initial or appended segment
        final String newSegmentVersion;
        if (versionOfExistingChunk != null) {
            // segment version overrides, so pick that now that we know it exists
            newSegmentVersion = versionOfExistingChunk;
        } else if (!pendings.isEmpty() && maxId != null) {
            // there is no visible segments in the time chunk, so pick the maxId of pendings, as computed above
            newSegmentVersion = maxId.getVersion();
        } else {
            // no segments, no pendings, so this must be the very first segment created for this interval
            newSegmentVersion = null;
        }
        if (maxId == null) {
            // When appending segments, null maxId means that we are allocating the very initial
            // segment for this time chunk.
            // This code is executed when the Overlord coordinates segment allocation, which is either you append segments
            // or you use segment lock. Since the core partitions set is not determined for appended segments, we set
            // it 0. When you use segment lock, the core partitions set doesn't work with it. We simply set it 0 so that the
            // OvershadowableManager handles the atomic segment update.
            final int newPartitionId = partialShardSpec.useNonRootGenerationPartitionSpace() ? PartitionIds.NON_ROOT_GEN_START_PARTITION_ID : PartitionIds.ROOT_GEN_START_PARTITION_ID;
            String version = newSegmentVersion == null ? existingVersion : newSegmentVersion;
            return new SegmentIdWithShardSpec(dataSource, interval, version, partialShardSpec.complete(jsonMapper, newPartitionId, 0));
        } else if (!maxId.getInterval().equals(interval) || maxId.getVersion().compareTo(existingVersion) > 0) {
            log.warn("Cannot allocate new segment for dataSource[%s], interval[%s], existingVersion[%s]: conflicting segment[%s].", dataSource, interval, existingVersion, maxId);
            return null;
        } else if (maxId.getShardSpec().getNumCorePartitions() == SingleDimensionShardSpec.UNKNOWN_NUM_CORE_PARTITIONS) {
            log.warn("Cannot allocate new segment because of unknown core partition size of segment[%s], shardSpec[%s]", maxId, maxId.getShardSpec());
            return null;
        } else {
            return new SegmentIdWithShardSpec(dataSource, maxId.getInterval(), Preconditions.checkNotNull(newSegmentVersion, "newSegmentVersion"), partialShardSpec.complete(jsonMapper, maxId.getShardSpec().getPartitionNum() + 1, maxId.getShardSpec().getNumCorePartitions()));
        }
    }
}