Package com.linkedin.davinci.kafka.consumer

Class LeaderFollowerStoreIngestionTask

java.lang.Object

com.linkedin.davinci.kafka.consumer.StoreIngestionTask

com.linkedin.davinci.kafka.consumer.LeaderFollowerStoreIngestionTask

All Implemented Interfaces:

java.io.Closeable, java.lang.AutoCloseable, java.lang.Runnable

Direct Known Subclasses:

ActiveActiveStoreIngestionTask

public class LeaderFollowerStoreIngestionTask
extends StoreIngestionTask

This class contains the state transition work between leader and follower; both leader and follower will keep track of information like which topic leader is consuming from and the corresponding offset as well as the latest successfully consumed or produced offset in the version topic (VT). State Transition: 1. OFFLINE -> STANDBY: Generate a SUBSCRIBE message in the consumer action queue; the logic here is the same as Online/Offline model; all it needs to do is to restore the checkpointed state from OffsetRecord; 2. STANDBY -> LEADER: The partition will be marked as in the transition progress from STANDBY to LEADER and completes the action immediately; after processing the rest of the consumer actions in the queue, check whether there is any partition is in the transition progress, if so: (i) consume the latest messages from version topic; (ii) drain all the messages in drainer queue in order to update the latest consumed message replication metadata; (iii) check whether there has been at least 5 minutes (configurable) of inactivity for this partition (meaning no new messages); if so, turn on the LEADER flag for this partition. 3. LEADER -> STANDBY: a. if the leader is consuming from VT, just set "isLeader" field to false and resume consumption; b. if the leader is consuming from anything other than VT, it needs to unsubscribe from the leader topic for this partition first, drain all the messages in the drainer queue for this leader topic/partition so that it can get the last producer callback for the last message it produces to VT; block on getting the result from the callback to update the corresponding offset in version topic, so that the new follower can subscribe back to VT using the recently updated VT offset.

Nested Class Summary

Nested classes/interfaces inherited from class com.linkedin.davinci.kafka.consumer.StoreIngestionTask

StoreIngestionTask.DelegateConsumerRecordResult

Field Summary

Fields

Modifier and Type	Field	Description
protected it.unimi.dsi.fastutil.ints.Int2ObjectMap<java.lang.String>	kafkaClusterIdToUrlMap
static java.util.function.Predicate<? super PartitionConsumptionState>	LEADER_OFFSET_LAG_FILTER
protected AvroStoreDeserializerCache	storeDeserializerCache
protected static java.util.function.LongPredicate	VALID_LAG
protected Lazy<VeniceWriter<byte[],byte[],byte[]>>	veniceWriter	N.B.: With L/F+native replication and many Leader partitions getting assigned to a single SN this VeniceWriter may be called from multiple thread simultaneously, during start of batch push.
protected java.util.Map<java.lang.String,VeniceViewWriter>	viewWriters

Fields inherited from class com.linkedin.davinci.kafka.consumer.StoreIngestionTask

aggKafkaConsumerService, amplificationFactor, amplificationFactorAdapter, availableSchemaIds, bootstrapTimeoutInMs, chunkAssembler, compressionStrategy, compressor, compressorFactory, consumerActionSequenceNumber, consumerActionsQueue, databaseSyncBytesIntervalForDeferredWriteMode, databaseSyncBytesIntervalForTransactionalMode, dataRecoverySourceVersionNumber, defaultReadyToServeChecker, deserializedSchemaIds, diskUsage, divErrorMetricCallback, emitMetrics, emptyPollSleepMs, errorPartitionId, hostLevelIngestionStats, hybridStoreConfig, idleCounter, ingestionTaskName, isChunked, isCurrentVersion, isDataRecovery, isDaVinciClient, isIsolatedIngestion, isRunning, isWriteComputationEnabled, kafkaClusterUrlResolver, kafkaProps, kafkaVersionTopic, KILL_WAIT_TIME_MS, localKafkaClusterId, localKafkaServer, localKafkaServerSingletonSet, metaStoreWriter, partitionConsumptionStateMap, partitionToPendingConsumerActionCountMap, pubSubTopicRepository, readCycleDelayMs, readOnlyForBatchOnlyStoreEnabled, realTimeTopic, recordLevelMetricEnabled, REDUNDANT_LOGGING_FILTER, SCHEMA_POLLING_DELAY_MS, schemaRepository, serverConfig, statusReportAdapter, storageEngine, storageEngineRepository, storageMetadataService, STORE_VERSION_POLLING_DELAY_MS, storeBufferService, storeName, storeRepository, storeVersionPartitionCount, subPartitionCount, topicManagerRepository, venicePartitioner, versionedDIVStats, versionedIngestionStats, versionNumber, versionTopic, writeComputeFailureCode

Constructor Summary

Constructors

Constructor

Description

LeaderFollowerStoreIngestionTask(StoreIngestionTaskFactory.Builder builder, Store store, Version version, java.util.Properties kafkaConsumerProperties, java.util.function.BooleanSupplier isCurrentVersion, VeniceStoreVersionConfig storeConfig, int errorPartitionId, boolean isIsolatedIngestion, java.util.Optional<ObjectCacheBackend> cacheBackend, java.util.function.Function<java.lang.Integer,DaVinciRecordTransformer> getRecordTransformer)

Method Summary

Modifier and Type	Method	Description
protected void	addPartitionConsumptionState(java.lang.Integer partition, PartitionConsumptionState pcs)
protected java.util.Map<java.lang.String,java.lang.Long>	calculateLeaderUpstreamOffsetWithTopicSwitch(PartitionConsumptionState partitionConsumptionState, TopicSwitch topicSwitch, PubSubTopic newSourceTopic, java.util.List<java.lang.CharSequence> unreachableBrokerList)
protected static void	checkAndHandleUpstreamOffsetRewind(PartitionConsumptionState partitionConsumptionState, PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord, long newUpstreamOffset, long previousUpstreamOffset, LeaderFollowerStoreIngestionTask ingestionTask)
protected boolean	checkAndLogIfLagIsAcceptableForHybridStore(PartitionConsumptionState pcs, long lag, long threshold, boolean shouldLogLag, LagType lagType, long latestConsumedProducerTimestamp)	Checks whether the lag is acceptable for hybrid stores
protected void	checkLongRunningTaskState()	The following function will be executed after processing all the quick actions in the consumer action queues, so that the long running actions doesn't block other partition's consumer actions.
protected void	closeVeniceViewWriters()
void	closeVeniceWriters(boolean doFlush)
void	consumerUnSubscribeAllTopics(PartitionConsumptionState partitionConsumptionState)	Unsubscribe from all the topics being consumed for the partition in partitionConsumptionState
protected LeaderProducerCallback	createProducerCallback(PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord, PartitionConsumptionState partitionConsumptionState, LeaderProducedRecordContext leaderProducedRecordContext, int subPartition, java.lang.String kafkaUrl, long beforeProcessingRecordTimestampNs)
protected StoreIngestionTask.DelegateConsumerRecordResult	delegateConsumerRecord(PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord, int subPartition, java.lang.String kafkaUrl, int kafkaClusterId, long beforeProcessingPerRecordTimestampNs, long beforeProcessingBatchRecordsTimestampMs)	The goal of this function is to possibly produce the incoming kafka message consumed from local VT, remote VT, RT or SR topic to local VT if needed.
void	demoteToStandby(PubSubTopicPartition topicPartition, LeaderFollowerPartitionStateModel.LeaderSessionIdChecker checker)
protected void	getAndUpdateLeaderCompletedState(KafkaKey kafkaKey, KafkaMessageEnvelope kafkaValue, ControlMessage controlMessage, PubSubMessageHeaders pubSubMessageHeaders, PartitionConsumptionState partitionConsumptionState)	HeartBeat SOS messages carry the leader completion state in the header.
long	getBatchFollowerOffsetLag()
long	getBatchLeaderOffsetLag()
long	getBatchReplicationLag()
protected java.util.Set<java.lang.String>	getConsumptionSourceKafkaAddress(PartitionConsumptionState partitionConsumptionState)
long	getFollowerOffsetLag()	Measure the offset lag between follower and leader
long	getHybridFollowerOffsetLag()
long	getHybridLeaderOffsetLag()
protected long	getLatestConsumedUpstreamOffsetForHybridOffsetLagMeasurement(PartitionConsumptionState pcs, java.lang.String ignoredKafkaUrl)	For regular L/F stores without A/A enabled, there is always only one real-time source.
protected long	getLatestPersistedUpstreamOffsetForHybridOffsetLagMeasurement(PartitionConsumptionState pcs, java.lang.String ignoredUpstreamKafkaUrl)	For L/F or NR, there is only one entry in upstreamOffsetMap whose key is NON_AA_REPLICATION_UPSTREAM_OFFSET_MAP_KEY.
long	getLeaderOffsetLag()
protected java.util.Set<java.lang.String>	getRealTimeDataSourceKafkaAddress(PartitionConsumptionState partitionConsumptionState)
long	getRegionHybridOffsetLag(int regionId)
protected int	getSubPartitionId(byte[] key, PubSubTopicPartition topicPartition)
protected long	getTopicPartitionOffsetByKafkaURL(java.lang.CharSequence kafkaURL, PubSubTopicPartition pubSubTopicPartition, long rewindStartTimestamp)
protected Lazy<VeniceWriter<byte[],byte[],byte[]>>	getVeniceWriter()
int	getWriteComputeErrorCode()
protected boolean	isHybridFollower(PartitionConsumptionState partitionConsumptionState)
protected boolean	isLeader(PartitionConsumptionState partitionConsumptionState)
boolean	isReadyToServeAnnouncedWithRTLag()
protected boolean	isRealTimeBufferReplayStarted(PartitionConsumptionState partitionConsumptionState)
protected void	leaderExecuteTopicSwitch(PartitionConsumptionState partitionConsumptionState, TopicSwitch topicSwitch, PubSubTopic newSourceTopic)
protected java.nio.ByteBuffer	maybeCompressData(int partition, java.nio.ByteBuffer data, PartitionConsumptionState partitionConsumptionState)	Compresses data in a bytebuffer when consuming from rt as a leader node and compression is enabled for the store version for which we're consuming data.
protected java.util.Set<java.lang.String>	maybeSendIngestionHeartbeat()	For hybrid stores only, the leader periodically writes a special SOS message to the RT topic with the following properties: 1.
protected long	measureHybridOffsetLag(PartitionConsumptionState partitionConsumptionState, boolean shouldLogLag)	For Leader/Follower state model, we already keep track of the consumption progress in leader, so directly calculate the lag with the real-time topic and the leader consumption offset.
protected long	measureRTOffsetLagForMultiRegions(java.util.Set<java.lang.String> sourceRealTimeTopicKafkaURLs, PartitionConsumptionState partitionConsumptionState, boolean shouldLogLag)
protected long	measureRTOffsetLagForSingleRegion(java.lang.String sourceRealTimeTopicKafkaURL, PartitionConsumptionState pcs, boolean shouldLog)	This method fetches/calculates latest leader persisted offset and last offset in RT topic.
protected void	processConsumerAction(ConsumerAction message, Store store)
protected void	processMessageAndMaybeProduceToKafka(PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord, PartitionConsumptionState partitionConsumptionState, int subPartition, java.lang.String kafkaUrl, int kafkaClusterId, long beforeProcessingRecordTimestampNs, long currentTimeForMetricsMs)
protected boolean	processTopicSwitch(ControlMessage controlMessage, int partition, long offset, PartitionConsumptionState partitionConsumptionState)	Process TopicSwitch control message at given partition offset for a specific PartitionConsumptionState.
protected void	processVersionSwapMessage(ControlMessage controlMessage, int partition, PartitionConsumptionState partitionConsumptionState)	This isn't really used for ingestion outside of A/A, so we NoOp here and rely on the actual implementation in ActiveActiveStoreIngestionTask
protected void	produceToLocalKafka(PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord, PartitionConsumptionState partitionConsumptionState, LeaderProducedRecordContext leaderProducedRecordContext, java.util.function.BiConsumer<ChunkAwareCallback,LeaderMetadataWrapper> produceFunction, int subPartition, java.lang.String kafkaUrl, int kafkaClusterId, long beforeProcessingRecordTimestampNs)
void	promoteToLeader(PubSubTopicPartition topicPartition, LeaderFollowerPartitionStateModel.LeaderSessionIdChecker checker)
protected void	recordHeartbeatReceived(PartitionConsumptionState partitionConsumptionState, PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord, java.lang.String kafkaUrl)
protected void	recordProcessedRecordStats(PartitionConsumptionState partitionConsumptionState, int processedRecordSize)
protected void	recordWriterStats(long consumerTimestampMs, long producerBrokerLatencyMs, long brokerConsumerLatencyMs, PartitionConsumptionState partitionConsumptionState)
protected void	reportIfCatchUpVersionTopicOffset(PartitionConsumptionState pcs)	Check if the ingestion progress has reached to the end of the version topic.
protected boolean	shouldCheckLeaderCompleteStateInFollower()	For non AA hybrid stores with AGGREGATE DRP, SIT reads from parent RT while the HB is written to the child RTs.
protected boolean	shouldCompressData(PartitionConsumptionState partitionConsumptionState)
protected boolean	shouldNewLeaderSwitchToRemoteConsumption(PartitionConsumptionState partitionConsumptionState)
protected boolean	shouldPersistRecord(PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> record, PartitionConsumptionState partitionConsumptionState)	Additional safeguards in Leader/Follower ingestion: 1.
protected boolean	shouldProcessRecord(PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> record, int subPartition)	For Leader/Follower model, the follower should have the same kind of check as the Online/Offline model; for leader, it's possible that it consumers from real-time topic or GF topic.
protected boolean	shouldProduceToVersionTopic(PartitionConsumptionState partitionConsumptionState)	For the corresponding partition being tracked in `partitionConsumptionState`, if it's in LEADER state and it's not consuming from version topic, it should produce the new message to version topic; besides, if LEADER is consuming remotely, it should also produce to local fabric.
protected void	startConsumingAsLeader(PartitionConsumptionState partitionConsumptionState)
protected void	startConsumingAsLeaderInTransitionFromStandby(PartitionConsumptionState partitionConsumptionState)
protected void	syncConsumedUpstreamRTOffsetMapIfNeeded(PartitionConsumptionState pcs, java.util.Map<java.lang.String,java.lang.Long> upstreamStartOffsetByKafkaURL)
protected void	syncTopicSwitchToIngestionMetadataService(TopicSwitch topicSwitch, PartitionConsumptionState partitionConsumptionState)
protected void	updateLatestInMemoryLeaderConsumedRTOffset(PartitionConsumptionState pcs, java.lang.String ignoredKafkaUrl, long offset)
protected void	updateLatestInMemoryProcessedOffset(PartitionConsumptionState partitionConsumptionState, PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecordWrapper, LeaderProducedRecordContext leaderProducedRecordContext, java.lang.String kafkaUrl, boolean dryRun)	Maintain the latest processed offsets by drainers in memory; in most of the time, these offsets are ahead of the checkpoint offsets inside OffsetRecord.
void	updateLeaderTopicOnFollower(PartitionConsumptionState partitionConsumptionState)
protected void	updateOffsetMetadataInOffsetRecord(PartitionConsumptionState partitionConsumptionState)	Sync the metadata about offset in OffsetRecord.
protected void	updateOffsetsFromConsumerRecord(PartitionConsumptionState partitionConsumptionState, PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord, LeaderProducedRecordContext leaderProducedRecordContext, com.linkedin.davinci.kafka.consumer.LeaderFollowerStoreIngestionTask.UpdateVersionTopicOffset updateVersionTopicOffsetFunction, com.linkedin.davinci.kafka.consumer.LeaderFollowerStoreIngestionTask.UpdateUpstreamTopicOffset updateUpstreamTopicOffsetFunction, com.linkedin.davinci.kafka.consumer.LeaderFollowerStoreIngestionTask.GetLastKnownUpstreamTopicOffset lastKnownUpstreamTopicOffsetSupplier, java.util.function.Supplier<java.lang.String> sourceKafkaUrlSupplier, boolean dryRun)	A helper function to the latest in-memory offsets processed by drainers in PartitionConsumptionState, after processing the given PubSubMessage.
protected void	waitForAllMessageToBeProcessedFromTopicPartition(PubSubTopicPartition topicPartition, PartitionConsumptionState partitionConsumptionState)	Besides draining messages in the drainer queue, wait for the last producer future.
protected void	waitForLastLeaderPersistFuture(PartitionConsumptionState partitionConsumptionState, java.lang.String errorMsg)

Methods inherited from class com.linkedin.davinci.kafka.consumer.StoreIngestionTask

checkIngestionProgress, cloneProducerStates, close, consumerBatchUnsubscribe, consumerHasAnySubscription, consumerHasSubscription, consumerResetOffset, consumerSubscribe, consumerUnSubscribe, createKafkaConsumerProperties, disableMetricsEmission, dumpPartitionConsumptionStates, dumpStoreVersionState, enableMetricsEmission, getAmplificationFactor, getCompressionStrategy, getCompressor, getFailedIngestionPartitionCount, getHostLevelIngestionStats, getIngestionTaskName, getKafkaVersionTopic, getLatestConsumedProducerTimestampWithSubPartition, getLeaderState, getOffsetToOnlineLagThresholdPerPartition, getPartitionConsumptionState, getPartitionOffsetLagBasedOnMetrics, getSchemaRepo, getServerConfig, getStatusReportAdapter, getStorageEngine, getStoreName, getTopicManager, getTopicPartitionEndOffSet, getVersionedDIVStats, getVersionIngestionStats, getVersionNumber, getVersionTopic, hasAnyPartitionConsumptionState, hasAnySubscription, hasPendingPartitionIngestionAction, isActiveActiveReplicationEnabled, isChunked, isFutureVersion, isHybridMode, isIngestionTaskActive, isMetricsEmissionEnabled, isPartitionConsumingOrHasPendingIngestionAction, isProducingVersionTopicHealthy, isReadyToServe, isRunning, isSegmentControlMsg, isStuckByMemoryConstraint, isTransientRecordBufferUsed, isUserSystemStore, kill, logStorageOperationWhileUnsubscribed, measureLagWithCallToPubSub, measureLagWithCallToPubSub, minZeroLag, nextSeqNum, processCommonConsumerAction, processConsumerRecord, processEndOfIncrementalPush, processEndOfPush, processStartOfIncrementalPush, produceToStoreBufferService, produceToStoreBufferServiceOrKafka, putInStorageEngine, recordChecksumVerificationFailure, removeFromStorageEngine, reportError, resetPartitionConsumptionOffset, resolveSourceKafkaServersWithinTopicSwitch, run, setLastConsumerException, setLastStoreIngestionException, setPartitionConsumptionState, shouldUpdateUpstreamOffset, shutdown, subscribePartition, subscribePartition, throwIfNotRunning, throwOrLogStorageFailureDependingIfStillSubscribed, unSubscribePartition, unSubscribePartition, updateOffsetMetadataAndSync, validateMessage, waitVersionStateAvailable

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

veniceWriter

protected final Lazy<VeniceWriter<byte[],byte[],byte[]>> veniceWriter

N.B.: With L/F+native replication and many Leader partitions getting assigned to a single SN this VeniceWriter may be called from multiple thread simultaneously, during start of batch push. Therefore, we wrap it in Lazy to initialize it in a thread safe way and to ensure that only one instance is created for the entire ingestion task.

kafkaClusterIdToUrlMap

protected final it.unimi.dsi.fastutil.ints.Int2ObjectMap<java.lang.String> kafkaClusterIdToUrlMap

viewWriters

protected final java.util.Map<java.lang.String,VeniceViewWriter> viewWriters

storeDeserializerCache

protected final AvroStoreDeserializerCache storeDeserializerCache

VALID_LAG

protected static final java.util.function.LongPredicate VALID_LAG

LEADER_OFFSET_LAG_FILTER

public static final java.util.function.Predicate<? super PartitionConsumptionState> LEADER_OFFSET_LAG_FILTER

Constructor Detail

LeaderFollowerStoreIngestionTask

public LeaderFollowerStoreIngestionTask(StoreIngestionTaskFactory.Builder builder,
                                        Store store,
                                        Version version,
                                        java.util.Properties kafkaConsumerProperties,
                                        java.util.function.BooleanSupplier isCurrentVersion,
                                        VeniceStoreVersionConfig storeConfig,
                                        int errorPartitionId,
                                        boolean isIsolatedIngestion,
                                        java.util.Optional<ObjectCacheBackend> cacheBackend,
                                        java.util.function.Function<java.lang.Integer,DaVinciRecordTransformer> getRecordTransformer)

Method Detail

closeVeniceWriters

public void closeVeniceWriters(boolean doFlush)

Overrides:

closeVeniceWriters in class StoreIngestionTask

closeVeniceViewWriters

protected void closeVeniceViewWriters()

Overrides:

closeVeniceViewWriters in class StoreIngestionTask

promoteToLeader

public void promoteToLeader(PubSubTopicPartition topicPartition,
                            LeaderFollowerPartitionStateModel.LeaderSessionIdChecker checker)

Specified by:

promoteToLeader in class StoreIngestionTask

demoteToStandby

public void demoteToStandby(PubSubTopicPartition topicPartition,
                            LeaderFollowerPartitionStateModel.LeaderSessionIdChecker checker)

Specified by:

demoteToStandby in class StoreIngestionTask

processConsumerAction

protected void processConsumerAction(ConsumerAction message,
                                     Store store)
                              throws java.lang.InterruptedException

Specified by:

processConsumerAction in class StoreIngestionTask

Throws:

java.lang.InterruptedException

checkLongRunningTaskState

protected void checkLongRunningTaskState()
                                  throws java.lang.InterruptedException

The following function will be executed after processing all the quick actions in the consumer action queues, so that the long running actions doesn't block other partition's consumer actions. Besides, there is no thread sleeping operations in this function in order to be efficient, but this function will be invoked again and again in the main loop of the StoreIngestionTask to check whether some long-running actions can finish now. The only drawback is that for regular batch push, leader flag is never on at least a few minutes after the leader consumes the last message (END_OF_PUSH), which is an acceptable trade-off for us in order to share and test the same code path between regular push job, hybrid store and reprocessing job.

Specified by:

checkLongRunningTaskState in class StoreIngestionTask

Throws:

java.lang.InterruptedException

startConsumingAsLeaderInTransitionFromStandby

protected void startConsumingAsLeaderInTransitionFromStandby(PartitionConsumptionState partitionConsumptionState)

calculateLeaderUpstreamOffsetWithTopicSwitch

protected java.util.Map<java.lang.String,java.lang.Long> calculateLeaderUpstreamOffsetWithTopicSwitch(PartitionConsumptionState partitionConsumptionState,
                                                                                                            TopicSwitch topicSwitch,
                                                                                                            PubSubTopic newSourceTopic,
                                                                                                            java.util.List<java.lang.CharSequence> unreachableBrokerList)

startConsumingAsLeader

protected void startConsumingAsLeader(PartitionConsumptionState partitionConsumptionState)

Overrides:

startConsumingAsLeader in class StoreIngestionTask

leaderExecuteTopicSwitch

protected void leaderExecuteTopicSwitch(PartitionConsumptionState partitionConsumptionState,
                                        TopicSwitch topicSwitch,
                                        PubSubTopic newSourceTopic)

syncConsumedUpstreamRTOffsetMapIfNeeded

protected void syncConsumedUpstreamRTOffsetMapIfNeeded(PartitionConsumptionState pcs,
                                                       java.util.Map<java.lang.String,java.lang.Long> upstreamStartOffsetByKafkaURL)

waitForLastLeaderPersistFuture

protected void waitForLastLeaderPersistFuture(PartitionConsumptionState partitionConsumptionState,
                                              java.lang.String errorMsg)

getTopicPartitionOffsetByKafkaURL

protected long getTopicPartitionOffsetByKafkaURL(java.lang.CharSequence kafkaURL,
                                                 PubSubTopicPartition pubSubTopicPartition,
                                                 long rewindStartTimestamp)

getConsumptionSourceKafkaAddress

protected java.util.Set<java.lang.String> getConsumptionSourceKafkaAddress(PartitionConsumptionState partitionConsumptionState)

Specified by:

getConsumptionSourceKafkaAddress in class StoreIngestionTask

getRealTimeDataSourceKafkaAddress

protected java.util.Set<java.lang.String> getRealTimeDataSourceKafkaAddress(PartitionConsumptionState partitionConsumptionState)

Overrides:

getRealTimeDataSourceKafkaAddress in class StoreIngestionTask

shouldNewLeaderSwitchToRemoteConsumption

protected boolean shouldNewLeaderSwitchToRemoteConsumption(PartitionConsumptionState partitionConsumptionState)

shouldProduceToVersionTopic

protected boolean shouldProduceToVersionTopic(PartitionConsumptionState partitionConsumptionState)

For the corresponding partition being tracked in `partitionConsumptionState`, if it's in LEADER state and it's not consuming from version topic, it should produce the new message to version topic; besides, if LEADER is consuming remotely, it should also produce to local fabric. If buffer replay is disable, all replicas will stick to version topic, no one is going to produce any message.

isLeader

protected boolean isLeader(PartitionConsumptionState partitionConsumptionState)

processTopicSwitch

protected boolean processTopicSwitch(ControlMessage controlMessage,
                                     int partition,
                                     long offset,
                                     PartitionConsumptionState partitionConsumptionState)

Process TopicSwitch control message at given partition offset for a specific PartitionConsumptionState. Return whether we need to execute additional ready-to-serve check after this message is processed.

Overrides:

processTopicSwitch in class StoreIngestionTask

syncTopicSwitchToIngestionMetadataService

protected void syncTopicSwitchToIngestionMetadataService(TopicSwitch topicSwitch,
                                                         PartitionConsumptionState partitionConsumptionState)

updateOffsetsFromConsumerRecord

protected void updateOffsetsFromConsumerRecord(PartitionConsumptionState partitionConsumptionState,
                                               PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord,
                                               LeaderProducedRecordContext leaderProducedRecordContext,
                                               com.linkedin.davinci.kafka.consumer.LeaderFollowerStoreIngestionTask.UpdateVersionTopicOffset updateVersionTopicOffsetFunction,
                                               com.linkedin.davinci.kafka.consumer.LeaderFollowerStoreIngestionTask.UpdateUpstreamTopicOffset updateUpstreamTopicOffsetFunction,
                                               com.linkedin.davinci.kafka.consumer.LeaderFollowerStoreIngestionTask.GetLastKnownUpstreamTopicOffset lastKnownUpstreamTopicOffsetSupplier,
                                               java.util.function.Supplier<java.lang.String> sourceKafkaUrlSupplier,
                                               boolean dryRun)

A helper function to the latest in-memory offsets processed by drainers in PartitionConsumptionState, after processing the given PubSubMessage. When using this helper function to update the latest in-memory offsets processed by drainers in PartitionConsumptionState: "updateVersionTopicOffsetFunction" should try to update the VT offset in PartitionConsumptionState "updateRealtimeTopicOffsetFunction" should try to update the latest processed upstream offset map in PartitionConsumptionState In LeaderFollowerStoreIngestionTask, "sourceKafkaUrlSupplier" should always return OffsetRecord.NON_AA_REPLICATION_UPSTREAM_OFFSET_MAP_KEY; in ActiveActiveStoreIngestionTask, "sourceKafkaUrlSupplier" should return the actual source Kafka url of the "consumerRecordWrapper" Dry-run mode would only check whether the offset rewind is benign or not instead of persisting the processed offset.

updateOffsetMetadataInOffsetRecord

protected void updateOffsetMetadataInOffsetRecord(PartitionConsumptionState partitionConsumptionState)

Description copied from class: StoreIngestionTask

Sync the metadata about offset in OffsetRecord. PartitionConsumptionState will pass through some information to OffsetRecord for persistence and Offset rewind/split brain has been guarded in StoreIngestionTask.updateLatestInMemoryProcessedOffset(com.linkedin.davinci.kafka.consumer.PartitionConsumptionState, com.linkedin.venice.pubsub.api.PubSubMessage<com.linkedin.venice.message.KafkaKey, com.linkedin.venice.kafka.protocol.KafkaMessageEnvelope, java.lang.Long>, com.linkedin.davinci.kafka.consumer.LeaderProducedRecordContext, java.lang.String, boolean).

Specified by:

updateOffsetMetadataInOffsetRecord in class StoreIngestionTask

updateLatestInMemoryProcessedOffset

protected void updateLatestInMemoryProcessedOffset(PartitionConsumptionState partitionConsumptionState,
                                                   PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecordWrapper,
                                                   LeaderProducedRecordContext leaderProducedRecordContext,
                                                   java.lang.String kafkaUrl,
                                                   boolean dryRun)

Description copied from class: StoreIngestionTask

Maintain the latest processed offsets by drainers in memory; in most of the time, these offsets are ahead of the checkpoint offsets inside OffsetRecord. Prior to update the offset in memory, the underlying storage engine should have persisted the given record. Dry-run mode will only do offset rewind check and it won't update the processed offset.

Specified by:

updateLatestInMemoryProcessedOffset in class StoreIngestionTask

checkAndHandleUpstreamOffsetRewind

protected static void checkAndHandleUpstreamOffsetRewind(PartitionConsumptionState partitionConsumptionState,
                                                         PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord,
                                                         long newUpstreamOffset,
                                                         long previousUpstreamOffset,
                                                         LeaderFollowerStoreIngestionTask ingestionTask)

produceToLocalKafka

protected void produceToLocalKafka(PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord,
                                   PartitionConsumptionState partitionConsumptionState,
                                   LeaderProducedRecordContext leaderProducedRecordContext,
                                   java.util.function.BiConsumer<ChunkAwareCallback,LeaderMetadataWrapper> produceFunction,
                                   int subPartition,
                                   java.lang.String kafkaUrl,
                                   int kafkaClusterId,
                                   long beforeProcessingRecordTimestampNs)

isRealTimeBufferReplayStarted

protected boolean isRealTimeBufferReplayStarted(PartitionConsumptionState partitionConsumptionState)

Specified by:

isRealTimeBufferReplayStarted in class StoreIngestionTask

measureHybridOffsetLag

protected long measureHybridOffsetLag(PartitionConsumptionState partitionConsumptionState,
                                      boolean shouldLogLag)

For Leader/Follower state model, we already keep track of the consumption progress in leader, so directly calculate the lag with the real-time topic and the leader consumption offset.

Specified by:

measureHybridOffsetLag in class StoreIngestionTask

measureRTOffsetLagForMultiRegions

protected long measureRTOffsetLagForMultiRegions(java.util.Set<java.lang.String> sourceRealTimeTopicKafkaURLs,
                                                 PartitionConsumptionState partitionConsumptionState,
                                                 boolean shouldLogLag)

isReadyToServeAnnouncedWithRTLag

public boolean isReadyToServeAnnouncedWithRTLag()

Overrides:

isReadyToServeAnnouncedWithRTLag in class StoreIngestionTask

reportIfCatchUpVersionTopicOffset

protected void reportIfCatchUpVersionTopicOffset(PartitionConsumptionState pcs)

Description copied from class: StoreIngestionTask

Check if the ingestion progress has reached to the end of the version topic. This is currently only used LeaderFollowerStoreIngestionTask.

Specified by:

reportIfCatchUpVersionTopicOffset in class StoreIngestionTask

shouldProcessRecord

protected boolean shouldProcessRecord(PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> record,
                                      int subPartition)

For Leader/Follower model, the follower should have the same kind of check as the Online/Offline model; for leader, it's possible that it consumers from real-time topic or GF topic.

Overrides:

shouldProcessRecord in class StoreIngestionTask

shouldPersistRecord

protected boolean shouldPersistRecord(PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> record,
                                      PartitionConsumptionState partitionConsumptionState)

Additional safeguards in Leader/Follower ingestion: 1. Check whether the incoming messages are from the expected source topics

Overrides:

shouldPersistRecord in class StoreIngestionTask

recordWriterStats

protected void recordWriterStats(long consumerTimestampMs,
                                 long producerBrokerLatencyMs,
                                 long brokerConsumerLatencyMs,
                                 PartitionConsumptionState partitionConsumptionState)

Overrides:

recordWriterStats in class StoreIngestionTask

recordProcessedRecordStats

protected void recordProcessedRecordStats(PartitionConsumptionState partitionConsumptionState,
                                          int processedRecordSize)

Overrides:

recordProcessedRecordStats in class StoreIngestionTask

isHybridFollower

protected boolean isHybridFollower(PartitionConsumptionState partitionConsumptionState)

Specified by:

isHybridFollower in class StoreIngestionTask

shouldCheckLeaderCompleteStateInFollower

protected boolean shouldCheckLeaderCompleteStateInFollower()

For non AA hybrid stores with AGGREGATE DRP, SIT reads from parent RT while the HB is written to the child RTs. Once all hybrid stores are either AA for cross colo replication and non AA otherwise, DRP and this extra check can also be removed.

Specified by:

shouldCheckLeaderCompleteStateInFollower in class StoreIngestionTask

checkAndLogIfLagIsAcceptableForHybridStore

protected boolean checkAndLogIfLagIsAcceptableForHybridStore(PartitionConsumptionState pcs,
                                                             long lag,
                                                             long threshold,
                                                             boolean shouldLogLag,
                                                             LagType lagType,
                                                             long latestConsumedProducerTimestamp)

Checks whether the lag is acceptable for hybrid stores

If the instance is a hybrid standby or DaVinciClient: Also check if
1. checkLeaderCompleteStateInFollower feature is enabled based on configs
2. leaderCompleteStatus has the leader state=completed and
3. the last update time was within the configured time interval to not use the stale leader state: check com.linkedin.venice.ConfigKeys.SERVER_LEADER_COMPLETE_STATE_CHECK_IN_FOLLOWER_VALID_INTERVAL_MS

Specified by:

checkAndLogIfLagIsAcceptableForHybridStore in class StoreIngestionTask

getAndUpdateLeaderCompletedState

protected void getAndUpdateLeaderCompletedState(KafkaKey kafkaKey,
                                                KafkaMessageEnvelope kafkaValue,
                                                ControlMessage controlMessage,
                                                PubSubMessageHeaders pubSubMessageHeaders,
                                                PartitionConsumptionState partitionConsumptionState)

HeartBeat SOS messages carry the leader completion state in the header. This function extracts the leader completion state from that header and updates the {@param partitionConsumptionState} accordingly.

recordHeartbeatReceived

protected void recordHeartbeatReceived(PartitionConsumptionState partitionConsumptionState,
                                       PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord,
                                       java.lang.String kafkaUrl)

Overrides:

recordHeartbeatReceived in class StoreIngestionTask

delegateConsumerRecord

protected StoreIngestionTask.DelegateConsumerRecordResult delegateConsumerRecord(PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord,
                                                                                 int subPartition,
                                                                                 java.lang.String kafkaUrl,
                                                                                 int kafkaClusterId,
                                                                                 long beforeProcessingPerRecordTimestampNs,
                                                                                 long beforeProcessingBatchRecordsTimestampMs)

The goal of this function is to possibly produce the incoming kafka message consumed from local VT, remote VT, RT or SR topic to local VT if needed. It's decided based on the function output of shouldProduceToVersionTopic(com.linkedin.davinci.kafka.consumer.PartitionConsumptionState) and message type. It also perform any necessary additional computation operation such as for write-compute/update message. It returns a boolean indicating if it was produced to kafka or not. This function should be called as one of the first steps in processing pipeline for all messages consumed from any kafka topic. The caller of this function should only process this {@param consumerRecord} further if the return is StoreIngestionTask.DelegateConsumerRecordResult.QUEUED_TO_DRAINER. This function assumes shouldProcessRecord(PubSubMessage, int) has been called which happens in StoreIngestionTask.produceToStoreBufferServiceOrKafka(Iterable, PubSubTopicPartition, String, int) before calling this and the it was decided that this record needs to be processed. It does not perform any validation check on the PartitionConsumptionState object to keep the goal of the function simple and not overload. Also DIV validation is done here if the message is received from RT topic. For more info please see please see StoreIngestionTask#internalProcessConsumerRecord(PubSubMessage, PartitionConsumptionState, LeaderProducedRecordContext, int, String, long) This function may modify the original record in KME and it is unsafe to use the payload from KME directly after this function.

Specified by:

delegateConsumerRecord in class StoreIngestionTask

Returns:

a StoreIngestionTask.DelegateConsumerRecordResult indicating what to do with the record

waitForAllMessageToBeProcessedFromTopicPartition

protected void waitForAllMessageToBeProcessedFromTopicPartition(PubSubTopicPartition topicPartition,
                                                                PartitionConsumptionState partitionConsumptionState)
                                                         throws java.lang.InterruptedException

Besides draining messages in the drainer queue, wait for the last producer future.

Overrides:

waitForAllMessageToBeProcessedFromTopicPartition in class StoreIngestionTask

Parameters:

topicPartition - for which to wait

Throws:

java.lang.InterruptedException

getBatchReplicationLag

public long getBatchReplicationLag()

Specified by:

getBatchReplicationLag in class StoreIngestionTask

getLeaderOffsetLag

public long getLeaderOffsetLag()

Specified by:

getLeaderOffsetLag in class StoreIngestionTask

getBatchLeaderOffsetLag

public long getBatchLeaderOffsetLag()

Specified by:

getBatchLeaderOffsetLag in class StoreIngestionTask

getHybridLeaderOffsetLag

public long getHybridLeaderOffsetLag()

Specified by:

getHybridLeaderOffsetLag in class StoreIngestionTask

getLatestPersistedUpstreamOffsetForHybridOffsetLagMeasurement

protected long getLatestPersistedUpstreamOffsetForHybridOffsetLagMeasurement(PartitionConsumptionState pcs,
                                                                             java.lang.String ignoredUpstreamKafkaUrl)

For L/F or NR, there is only one entry in upstreamOffsetMap whose key is NON_AA_REPLICATION_UPSTREAM_OFFSET_MAP_KEY. Return the value of the entry.

getLatestConsumedUpstreamOffsetForHybridOffsetLagMeasurement

protected long getLatestConsumedUpstreamOffsetForHybridOffsetLagMeasurement(PartitionConsumptionState pcs,
                                                                            java.lang.String ignoredKafkaUrl)

For regular L/F stores without A/A enabled, there is always only one real-time source.

updateLatestInMemoryLeaderConsumedRTOffset

protected void updateLatestInMemoryLeaderConsumedRTOffset(PartitionConsumptionState pcs,
                                                          java.lang.String ignoredKafkaUrl,
                                                          long offset)

getFollowerOffsetLag

public long getFollowerOffsetLag()

Description copied from class: StoreIngestionTask

Measure the offset lag between follower and leader

Specified by:

getFollowerOffsetLag in class StoreIngestionTask

getBatchFollowerOffsetLag

public long getBatchFollowerOffsetLag()

Specified by:

getBatchFollowerOffsetLag in class StoreIngestionTask

getHybridFollowerOffsetLag

public long getHybridFollowerOffsetLag()

Specified by:

getHybridFollowerOffsetLag in class StoreIngestionTask

getRegionHybridOffsetLag

public long getRegionHybridOffsetLag(int regionId)

Specified by:

getRegionHybridOffsetLag in class StoreIngestionTask

consumerUnSubscribeAllTopics

public void consumerUnSubscribeAllTopics(PartitionConsumptionState partitionConsumptionState)

Unsubscribe from all the topics being consumed for the partition in partitionConsumptionState

Specified by:

consumerUnSubscribeAllTopics in class StoreIngestionTask

getWriteComputeErrorCode

public int getWriteComputeErrorCode()

Specified by:

getWriteComputeErrorCode in class StoreIngestionTask

updateLeaderTopicOnFollower

public void updateLeaderTopicOnFollower(PartitionConsumptionState partitionConsumptionState)

Specified by:

updateLeaderTopicOnFollower in class StoreIngestionTask

getSubPartitionId

protected int getSubPartitionId(byte[] key,
                                PubSubTopicPartition topicPartition)

maybeCompressData

protected java.nio.ByteBuffer maybeCompressData(int partition,
                                                java.nio.ByteBuffer data,
                                                PartitionConsumptionState partitionConsumptionState)

Compresses data in a bytebuffer when consuming from rt as a leader node and compression is enabled for the store version for which we're consuming data.

Parameters:

partition - which partition we're acting on so as to determine the PartitionConsumptionState

data - the data that we might compress

Returns:

a bytebuffer that's either the original bytebuffer or a new one depending on if we compressed it.

shouldCompressData

protected boolean shouldCompressData(PartitionConsumptionState partitionConsumptionState)

processMessageAndMaybeProduceToKafka

protected void processMessageAndMaybeProduceToKafka(PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord,
                                                    PartitionConsumptionState partitionConsumptionState,
                                                    int subPartition,
                                                    java.lang.String kafkaUrl,
                                                    int kafkaClusterId,
                                                    long beforeProcessingRecordTimestampNs,
                                                    long currentTimeForMetricsMs)

measureRTOffsetLagForSingleRegion

protected long measureRTOffsetLagForSingleRegion(java.lang.String sourceRealTimeTopicKafkaURL,
                                                 PartitionConsumptionState pcs,
                                                 boolean shouldLog)

This method fetches/calculates latest leader persisted offset and last offset in RT topic. The method relies on getLatestPersistedUpstreamOffsetForHybridOffsetLagMeasurement(PartitionConsumptionState, String) to fetch latest leader persisted offset for different data replication policy.

Returns:

the lag (lastOffsetInRealTimeTopic - latestPersistedLeaderOffset)

processVersionSwapMessage

protected void processVersionSwapMessage(ControlMessage controlMessage,
                                         int partition,
                                         PartitionConsumptionState partitionConsumptionState)

Description copied from class: StoreIngestionTask

This isn't really used for ingestion outside of A/A, so we NoOp here and rely on the actual implementation in ActiveActiveStoreIngestionTask

Overrides:

processVersionSwapMessage in class StoreIngestionTask

createProducerCallback

protected LeaderProducerCallback createProducerCallback(PubSubMessage<KafkaKey,KafkaMessageEnvelope,java.lang.Long> consumerRecord,
                                                        PartitionConsumptionState partitionConsumptionState,
                                                        LeaderProducedRecordContext leaderProducedRecordContext,
                                                        int subPartition,
                                                        java.lang.String kafkaUrl,
                                                        long beforeProcessingRecordTimestampNs)

getVeniceWriter

protected Lazy<VeniceWriter<byte[],byte[],byte[]>> getVeniceWriter()

addPartitionConsumptionState

protected void addPartitionConsumptionState(java.lang.Integer partition,
                                            PartitionConsumptionState pcs)

maybeSendIngestionHeartbeat

protected java.util.Set<java.lang.String> maybeSendIngestionHeartbeat()

For hybrid stores only, the leader periodically writes a special SOS message to the RT topic with the following properties:
1. Special key: This key contains constant bytes, allowing for compaction.
2. Fixed/known producer GUID: This GUID is dedicated to heartbeats and prevents DIV from breaking.
3. Special segment: This segment never contains data, eliminating the need for an EOS message.

Upon consuming the SOS message, the leader writes it to its local VT. Once the drainer processes the record, the leader updates its latest processed upstream RT topic offset. At this point, the offset reflects the correct position, regardless of trailing CMs or skippable data records due to DCR.

This heartbeat message does not send a leader completion header. This results in having the leader completion states only in VTs and not in the RT, avoiding the need to differentiate between heartbeats from leaders of different versions (backup/current/future) and colos.

Specified by:

maybeSendIngestionHeartbeat in class StoreIngestionTask

Returns:

the set of partitions that failed to send heartbeat (used for tests)