use of org.apache.druid.server.coordinator.ServerHolder in project druid by druid-io.
the class BroadcastDistributionRuleTest method setUp.
@Before
public void setUp() {
smallSegment = new DataSegment("small_source", Intervals.of("0/1000"), DateTimes.nowUtc().toString(), new HashMap<>(), new ArrayList<>(), new ArrayList<>(), NoneShardSpec.instance(), 0, 0);
for (int i = 0; i < 3; i++) {
largeSegments.add(new DataSegment("large_source", Intervals.of((i * 1000) + "/" + ((i + 1) * 1000)), DateTimes.nowUtc().toString(), new HashMap<>(), new ArrayList<>(), new ArrayList<>(), NoneShardSpec.instance(), 0, 100));
}
for (int i = 0; i < 2; i++) {
largeSegments2.add(new DataSegment("large_source2", Intervals.of((i * 1000) + "/" + ((i + 1) * 1000)), DateTimes.nowUtc().toString(), new HashMap<>(), new ArrayList<>(), new ArrayList<>(), NoneShardSpec.instance(), 0, 100));
}
holderOfSmallSegment = new ServerHolder(new DruidServer("serverHot2", "hostHot2", null, 1000, ServerType.HISTORICAL, "hot", 0).addDataSegment(smallSegment).toImmutableDruidServer(), new LoadQueuePeonTester());
holdersOfLargeSegments.add(new ServerHolder(new DruidServer("serverHot1", "hostHot1", null, 1000, ServerType.HISTORICAL, "hot", 0).addDataSegment(largeSegments.get(0)).toImmutableDruidServer(), new LoadQueuePeonTester()));
holdersOfLargeSegments.add(new ServerHolder(new DruidServer("serverNorm1", "hostNorm1", null, 1000, ServerType.HISTORICAL, DruidServer.DEFAULT_TIER, 0).addDataSegment(largeSegments.get(1)).toImmutableDruidServer(), new LoadQueuePeonTester()));
holdersOfLargeSegments.add(new ServerHolder(new DruidServer("serverNorm2", "hostNorm2", null, 100, ServerType.HISTORICAL, DruidServer.DEFAULT_TIER, 0).addDataSegment(largeSegments.get(2)).toImmutableDruidServer(), new LoadQueuePeonTester()));
holdersOfLargeSegments2.add(new ServerHolder(new DruidServer("serverHot3", "hostHot3", null, 1000, ServerType.HISTORICAL, "hot", 0).addDataSegment(largeSegments2.get(0)).toImmutableDruidServer(), new LoadQueuePeonTester()));
holdersOfLargeSegments2.add(new ServerHolder(new DruidServer("serverNorm3", "hostNorm3", null, 100, ServerType.HISTORICAL, DruidServer.DEFAULT_TIER, 0).addDataSegment(largeSegments2.get(1)).toImmutableDruidServer(), new LoadQueuePeonTester()));
activeServer = new ServerHolder(new DruidServer("active", "host1", null, 100, ServerType.HISTORICAL, "tier1", 0).addDataSegment(largeSegments.get(0)).toImmutableDruidServer(), new LoadQueuePeonTester());
decommissioningServer1 = new ServerHolder(new DruidServer("decommissioning1", "host2", null, 100, ServerType.HISTORICAL, "tier1", 0).addDataSegment(smallSegment).toImmutableDruidServer(), new LoadQueuePeonTester(), true);
decommissioningServer2 = new ServerHolder(new DruidServer("decommissioning2", "host3", null, 100, ServerType.HISTORICAL, "tier1", 0).addDataSegment(largeSegments.get(1)).toImmutableDruidServer(), new LoadQueuePeonTester(), true);
druidCluster = DruidClusterBuilder.newBuilder().addTier("hot", holdersOfLargeSegments.get(0), holderOfSmallSegment, holdersOfLargeSegments2.get(0)).addTier(DruidServer.DEFAULT_TIER, holdersOfLargeSegments.get(1), holdersOfLargeSegments.get(2), holdersOfLargeSegments2.get(1)).build();
secondCluster = DruidClusterBuilder.newBuilder().addTier("tier1", activeServer, decommissioningServer1, decommissioningServer2).build();
}
use of org.apache.druid.server.coordinator.ServerHolder in project druid by druid-io.
the class BalanceSegments method balanceServers.
private Pair<Integer, Integer> balanceServers(DruidCoordinatorRuntimeParams params, List<ServerHolder> toMoveFrom, List<ServerHolder> toMoveTo, int maxSegmentsToMove) {
if (maxSegmentsToMove <= 0) {
log.debug("maxSegmentsToMove is 0; no balancing work can be performed.");
return new Pair<>(0, 0);
} else if (toMoveFrom.isEmpty()) {
log.debug("toMoveFrom is empty; no balancing work can be performed.");
return new Pair<>(0, 0);
} else if (toMoveTo.isEmpty()) {
log.debug("toMoveTo is empty; no balancing work can be peformed.");
return new Pair<>(0, 0);
}
final BalancerStrategy strategy = params.getBalancerStrategy();
final int maxIterations = 2 * maxSegmentsToMove;
final int maxToLoad = params.getCoordinatorDynamicConfig().getMaxSegmentsInNodeLoadingQueue();
int moved = 0, unmoved = 0;
Iterator<BalancerSegmentHolder> segmentsToMove;
// The pick method depends on if the operator has enabled batched segment sampling in the Coorinator dynamic config.
if (params.getCoordinatorDynamicConfig().useBatchedSegmentSampler()) {
segmentsToMove = strategy.pickSegmentsToMove(toMoveFrom, params.getBroadcastDatasources(), maxSegmentsToMove);
} else {
segmentsToMove = strategy.pickSegmentsToMove(toMoveFrom, params.getBroadcastDatasources(), params.getCoordinatorDynamicConfig().getPercentOfSegmentsToConsiderPerMove());
}
// noinspection ForLoopThatDoesntUseLoopVariable
for (int iter = 0; (moved + unmoved) < maxSegmentsToMove; ++iter) {
if (!segmentsToMove.hasNext()) {
log.info("All servers to move segments from are empty, ending run.");
break;
}
final BalancerSegmentHolder segmentToMoveHolder = segmentsToMove.next();
// DruidCoordinatorRuntimeParams.getUsedSegments originate from SegmentsMetadataManager, i. e. that's a set of segments
// that *should* be loaded. segmentToMoveHolder.getSegment originates from ServerInventoryView, i. e. that may be
// any segment that happens to be loaded on some server, even if it is not used. (Coordinator closes such
// discrepancies eventually via UnloadUnusedSegments). Therefore the picked segmentToMoveHolder's segment may not
// need to be balanced.
boolean needToBalancePickedSegment = params.getUsedSegments().contains(segmentToMoveHolder.getSegment());
if (needToBalancePickedSegment) {
final DataSegment segmentToMove = segmentToMoveHolder.getSegment();
final ImmutableDruidServer fromServer = segmentToMoveHolder.getFromServer();
// we want to leave the server the segment is currently on in the list...
// but filter out replicas that are already serving the segment, and servers with a full load queue
final List<ServerHolder> toMoveToWithLoadQueueCapacityAndNotServingSegment = toMoveTo.stream().filter(s -> s.getServer().equals(fromServer) || (!s.isServingSegment(segmentToMove) && (maxToLoad <= 0 || s.getNumberOfSegmentsInQueue() < maxToLoad))).collect(Collectors.toList());
if (toMoveToWithLoadQueueCapacityAndNotServingSegment.size() > 0) {
final ServerHolder destinationHolder = strategy.findNewSegmentHomeBalancer(segmentToMove, toMoveToWithLoadQueueCapacityAndNotServingSegment);
if (destinationHolder != null && !destinationHolder.getServer().equals(fromServer)) {
if (moveSegment(segmentToMoveHolder, destinationHolder.getServer(), params)) {
moved++;
} else {
unmoved++;
}
} else {
log.debug("Segment [%s] is 'optimally' placed.", segmentToMove.getId());
unmoved++;
}
} else {
log.debug("No valid movement destinations for segment [%s].", segmentToMove.getId());
unmoved++;
}
}
if (iter >= maxIterations) {
log.info("Unable to select %d remaining candidate segments out of %d total to balance " + "after %d iterations, ending run.", (maxSegmentsToMove - moved - unmoved), maxSegmentsToMove, iter);
break;
}
}
return new Pair<>(moved, unmoved);
}
use of org.apache.druid.server.coordinator.ServerHolder in project druid by druid-io.
the class EmitClusterStatsAndMetrics method run.
@Override
public DruidCoordinatorRuntimeParams run(DruidCoordinatorRuntimeParams params) {
DruidCluster cluster = params.getDruidCluster();
CoordinatorStats stats = params.getCoordinatorStats();
ServiceEmitter emitter = params.getEmitter();
stats.forEachTieredStat("assignedCount", (final String tier, final long count) -> {
log.info("[%s] : Assigned %s segments among %,d servers", tier, count, cluster.getHistoricalsByTier(tier).size());
emitTieredStat(emitter, "segment/assigned/count", tier, count);
});
stats.forEachTieredStat("droppedCount", (final String tier, final long count) -> {
log.info("[%s] : Dropped %s segments among %,d servers", tier, count, cluster.getHistoricalsByTier(tier).size());
emitTieredStat(emitter, "segment/dropped/count", tier, count);
});
emitTieredStats(emitter, "segment/cost/raw", stats, "initialCost");
emitTieredStats(emitter, "segment/cost/normalization", stats, "normalization");
emitTieredStats(emitter, "segment/moved/count", stats, "movedCount");
emitTieredStats(emitter, "segment/deleted/count", stats, "deletedCount");
stats.forEachTieredStat("normalizedInitialCostTimesOneThousand", (final String tier, final long count) -> {
emitTieredStat(emitter, "segment/cost/normalized", tier, count / 1000d);
});
stats.forEachTieredStat("unneededCount", (final String tier, final long count) -> {
log.info("[%s] : Removed %s unneeded segments among %,d servers", tier, count, cluster.getHistoricalsByTier(tier).size());
emitTieredStat(emitter, "segment/unneeded/count", tier, count);
});
emitter.emit(new ServiceMetricEvent.Builder().build("segment/overShadowed/count", stats.getGlobalStat("overShadowedCount")));
stats.forEachTieredStat("movedCount", (final String tier, final long count) -> {
log.info("[%s] : Moved %,d segment(s)", tier, count);
});
stats.forEachTieredStat("unmovedCount", (final String tier, final long count) -> {
log.info("[%s] : Let alone %,d segment(s)", tier, count);
});
log.info("Load Queues:");
for (Iterable<ServerHolder> serverHolders : cluster.getSortedHistoricalsByTier()) {
for (ServerHolder serverHolder : serverHolders) {
ImmutableDruidServer server = serverHolder.getServer();
LoadQueuePeon queuePeon = serverHolder.getPeon();
log.info("Server[%s, %s, %s] has %,d left to load, %,d left to drop, %,d bytes queued, %,d bytes served.", server.getName(), server.getType().toString(), server.getTier(), queuePeon.getSegmentsToLoad().size(), queuePeon.getSegmentsToDrop().size(), queuePeon.getLoadQueueSize(), server.getCurrSize());
if (log.isDebugEnabled()) {
for (DataSegment segment : queuePeon.getSegmentsToLoad()) {
log.debug("Segment to load[%s]", segment);
}
for (DataSegment segment : queuePeon.getSegmentsToDrop()) {
log.debug("Segment to drop[%s]", segment);
}
}
stats.addToTieredStat(TOTAL_CAPACITY, server.getTier(), server.getMaxSize());
stats.addToTieredStat(TOTAL_HISTORICAL_COUNT, server.getTier(), 1);
}
}
params.getDatabaseRuleManager().getAllRules().values().forEach(rules -> rules.forEach(rule -> {
if (rule instanceof LoadRule) {
((LoadRule) rule).getTieredReplicants().forEach((tier, replica) -> stats.accumulateMaxTieredStat(MAX_REPLICATION_FACTOR, tier, replica));
}
}));
emitTieredStats(emitter, "tier/required/capacity", stats, LoadRule.REQUIRED_CAPACITY);
emitTieredStats(emitter, "tier/total/capacity", stats, TOTAL_CAPACITY);
emitTieredStats(emitter, "tier/replication/factor", stats, MAX_REPLICATION_FACTOR);
emitTieredStats(emitter, "tier/historical/count", stats, TOTAL_HISTORICAL_COUNT);
// Emit coordinator metrics
params.getLoadManagementPeons().forEach((final String serverName, final LoadQueuePeon queuePeon) -> {
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.SERVER, serverName).build("segment/loadQueue/size", queuePeon.getLoadQueueSize()));
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.SERVER, serverName).build("segment/loadQueue/failed", queuePeon.getAndResetFailedAssignCount()));
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.SERVER, serverName).build("segment/loadQueue/count", queuePeon.getSegmentsToLoad().size()));
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.SERVER, serverName).build("segment/dropQueue/count", queuePeon.getSegmentsToDrop().size()));
});
coordinator.computeNumsUnavailableUsedSegmentsPerDataSource().object2IntEntrySet().forEach((final Object2IntMap.Entry<String> entry) -> {
final String dataSource = entry.getKey();
final int numUnavailableUsedSegmentsInDataSource = entry.getIntValue();
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.DATASOURCE, dataSource).build("segment/unavailable/count", numUnavailableUsedSegmentsInDataSource));
});
coordinator.computeUnderReplicationCountsPerDataSourcePerTier().forEach((final String tier, final Object2LongMap<String> underReplicationCountsPerDataSource) -> {
for (final Object2LongMap.Entry<String> entry : underReplicationCountsPerDataSource.object2LongEntrySet()) {
final String dataSource = entry.getKey();
final long underReplicationCount = entry.getLongValue();
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.TIER, tier).setDimension(DruidMetrics.DATASOURCE, dataSource).build("segment/underReplicated/count", underReplicationCount));
}
});
emitter.emit(new ServiceMetricEvent.Builder().build("compact/task/count", stats.getGlobalStat(CompactSegments.COMPACTION_TASK_COUNT)));
emitter.emit(new ServiceMetricEvent.Builder().build("compactTask/maxSlot/count", stats.getGlobalStat(CompactSegments.MAX_COMPACTION_TASK_SLOT)));
emitter.emit(new ServiceMetricEvent.Builder().build("compactTask/availableSlot/count", stats.getGlobalStat(CompactSegments.AVAILABLE_COMPACTION_TASK_SLOT)));
stats.forEachDataSourceStat(CompactSegments.TOTAL_SIZE_OF_SEGMENTS_AWAITING, (final String dataSource, final long count) -> {
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.DATASOURCE, dataSource).build("segment/waitCompact/bytes", count));
});
stats.forEachDataSourceStat(CompactSegments.TOTAL_COUNT_OF_SEGMENTS_AWAITING, (final String dataSource, final long count) -> {
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.DATASOURCE, dataSource).build("segment/waitCompact/count", count));
});
stats.forEachDataSourceStat(CompactSegments.TOTAL_INTERVAL_OF_SEGMENTS_AWAITING, (final String dataSource, final long count) -> {
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.DATASOURCE, dataSource).build("interval/waitCompact/count", count));
});
stats.forEachDataSourceStat(CompactSegments.TOTAL_SIZE_OF_SEGMENTS_SKIPPED, (final String dataSource, final long count) -> {
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.DATASOURCE, dataSource).build("segment/skipCompact/bytes", count));
});
stats.forEachDataSourceStat(CompactSegments.TOTAL_COUNT_OF_SEGMENTS_SKIPPED, (final String dataSource, final long count) -> {
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.DATASOURCE, dataSource).build("segment/skipCompact/count", count));
});
stats.forEachDataSourceStat(CompactSegments.TOTAL_INTERVAL_OF_SEGMENTS_SKIPPED, (final String dataSource, final long count) -> {
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.DATASOURCE, dataSource).build("interval/skipCompact/count", count));
});
stats.forEachDataSourceStat(CompactSegments.TOTAL_SIZE_OF_SEGMENTS_COMPACTED, (final String dataSource, final long count) -> {
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.DATASOURCE, dataSource).build("segment/compacted/bytes", count));
});
stats.forEachDataSourceStat(CompactSegments.TOTAL_COUNT_OF_SEGMENTS_COMPACTED, (final String dataSource, final long count) -> {
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.DATASOURCE, dataSource).build("segment/compacted/count", count));
});
stats.forEachDataSourceStat(CompactSegments.TOTAL_INTERVAL_OF_SEGMENTS_COMPACTED, (final String dataSource, final long count) -> {
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.DATASOURCE, dataSource).build("interval/compacted/count", count));
});
// Emit segment metrics
params.getUsedSegmentsTimelinesPerDataSource().forEach((String dataSource, VersionedIntervalTimeline<String, DataSegment> dataSourceWithUsedSegments) -> {
long totalSizeOfUsedSegments = dataSourceWithUsedSegments.iterateAllObjects().stream().mapToLong(DataSegment::getSize).sum();
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.DATASOURCE, dataSource).build("segment/size", totalSizeOfUsedSegments));
emitter.emit(new ServiceMetricEvent.Builder().setDimension(DruidMetrics.DATASOURCE, dataSource).build("segment/count", dataSourceWithUsedSegments.getNumObjects()));
});
// Emit coordinator runtime stats
emitDutyStats(emitter, "coordinator/time", stats, "runtime");
return params;
}
use of org.apache.druid.server.coordinator.ServerHolder in project druid by druid-io.
the class BalanceSegments method moveSegment.
protected boolean moveSegment(final BalancerSegmentHolder segment, final ImmutableDruidServer toServer, final DruidCoordinatorRuntimeParams params) {
final LoadQueuePeon toPeon = params.getLoadManagementPeons().get(toServer.getName());
final ImmutableDruidServer fromServer = segment.getFromServer();
final DataSegment segmentToMove = segment.getSegment();
final SegmentId segmentId = segmentToMove.getId();
if (!toPeon.getSegmentsToLoad().contains(segmentToMove) && (toServer.getSegment(segmentId) == null) && new ServerHolder(toServer, toPeon).getAvailableSize() > segmentToMove.getSize()) {
log.debug("Moving [%s] from [%s] to [%s]", segmentId, fromServer.getName(), toServer.getName());
LoadPeonCallback callback = null;
try {
ConcurrentMap<SegmentId, BalancerSegmentHolder> movingSegments = currentlyMovingSegments.get(toServer.getTier());
movingSegments.put(segmentId, segment);
callback = () -> movingSegments.remove(segmentId);
coordinator.moveSegment(params, fromServer, toServer, segmentToMove, callback);
return true;
} catch (Exception e) {
log.makeAlert(e, StringUtils.format("[%s] : Moving exception", segmentId)).emit();
if (callback != null) {
callback.execute();
}
}
}
return false;
}
use of org.apache.druid.server.coordinator.ServerHolder in project druid by druid-io.
the class UnloadUnusedSegments method run.
@Override
public DruidCoordinatorRuntimeParams run(DruidCoordinatorRuntimeParams params) {
CoordinatorStats stats = new CoordinatorStats();
Set<DataSegment> usedSegments = params.getUsedSegments();
DruidCluster cluster = params.getDruidCluster();
Map<String, Boolean> broadcastStatusByDatasource = new HashMap<>();
for (String broadcastDatasource : params.getBroadcastDatasources()) {
broadcastStatusByDatasource.put(broadcastDatasource, true);
}
for (SortedSet<ServerHolder> serverHolders : cluster.getSortedHistoricalsByTier()) {
for (ServerHolder serverHolder : serverHolders) {
handleUnusedSegmentsForServer(serverHolder, usedSegments, params, stats, false, broadcastStatusByDatasource);
}
}
for (ServerHolder serverHolder : cluster.getBrokers()) {
handleUnusedSegmentsForServer(serverHolder, usedSegments, params, stats, false, broadcastStatusByDatasource);
}
for (ServerHolder serverHolder : cluster.getRealtimes()) {
handleUnusedSegmentsForServer(serverHolder, usedSegments, params, stats, true, broadcastStatusByDatasource);
}
return params.buildFromExisting().withCoordinatorStats(stats).build();
}
Aggregations