Github user xccui commented on a diff in the pull request:

    https://github.com/apache/flink/pull/4625#discussion_r137201317
  
    --- Diff: 
flink-libraries/flink-table/src/main/scala/org/apache/flink/table/runtime/join/TimeBoundedStreamInnerJoin.scala
 ---
    @@ -0,0 +1,533 @@
    +/*
    + * Licensed to the Apache Software Foundation (ASF) under one
    + * or more contributor license agreements.  See the NOTICE file
    + * distributed with this work for additional information
    + * regarding copyright ownership.  The ASF licenses this file
    + * to you under the Apache License, Version 2.0 (the
    + * "License"); you may not use this file except in compliance
    + * with the License.  You may obtain a copy of the License at
    + *
    + *     http://www.apache.org/licenses/LICENSE-2.0
    + *
    + * Unless required by applicable law or agreed to in writing, software
    + * distributed under the License is distributed on an "AS IS" BASIS,
    + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    + * See the License for the specific language governing permissions and
    + * limitations under the License.
    + */
    +
    +package org.apache.flink.table.runtime.join
    +
    +import java.text.SimpleDateFormat
    +import java.util
    +import java.util.Map.Entry
    +import java.util.{Date, List => JList}
    +
    +import org.apache.flink.api.common.functions.FlatJoinFunction
    +import org.apache.flink.api.common.state._
    +import org.apache.flink.api.common.typeinfo.{BasicTypeInfo, 
TypeInformation}
    +import org.apache.flink.api.java.typeutils.ListTypeInfo
    +import org.apache.flink.configuration.Configuration
    +import org.apache.flink.streaming.api.functions.co.CoProcessFunction
    +import org.apache.flink.table.codegen.Compiler
    +import org.apache.flink.table.runtime.CRowWrappingCollector
    +import 
org.apache.flink.table.runtime.join.JoinTimeIndicator.JoinTimeIndicator
    +import org.apache.flink.table.runtime.types.CRow
    +import org.apache.flink.table.util.Logging
    +import org.apache.flink.types.Row
    +import org.apache.flink.util.Collector
    +
    +/**
    +  * A CoProcessFunction to execute time-bounded stream inner-join.
    +  *
    +  * Sample criteria:
    +  *
    +  * L.time between R.time + X and R.time + Y
    +  * or AND R.time between L.time - Y and L.time - X
    +  *
    +  * @param leftLowerBound  X
    +  * @param leftUpperBound  Y
    +  * @param allowedLateness the lateness allowed for the two streams
    +  * @param leftType        the input type of left stream
    +  * @param rightType       the input type of right stream
    +  * @param genJoinFuncName the function code of other non-equi conditions
    +  * @param genJoinFuncCode the function name of other non-equi conditions
    +  * @param timeIndicator   indicate whether joining on proctime or rowtime
    +  *
    +  */
    +class TimeBoundedStreamInnerJoin(
    +  private val leftLowerBound: Long,
    +  private val leftUpperBound: Long,
    +  private val allowedLateness: Long,
    +  private val leftType: TypeInformation[Row],
    +  private val rightType: TypeInformation[Row],
    +  private val genJoinFuncName: String,
    +  private val genJoinFuncCode: String,
    +  private val leftTimeIdx: Int,
    +  private val rightTimeIdx: Int,
    +  private val timeIndicator: JoinTimeIndicator)
    +  extends CoProcessFunction[CRow, CRow, CRow]
    +    with Compiler[FlatJoinFunction[Row, Row, Row]]
    +    with Logging {
    +
    +  private var cRowWrapper: CRowWrappingCollector = _
    +
    +  // the join function for other conditions
    +  private var joinFunction: FlatJoinFunction[Row, Row, Row] = _
    +
    +  // cache to store the left stream records
    +  private var leftCache: MapState[Long, JList[Row]] = _
    +  // cache to store right stream records
    +  private var rightCache: MapState[Long, JList[Row]] = _
    +
    +  // state to record the timer on the left stream. 0 means no timer set
    +  private var leftTimerState: ValueState[Long] = _
    +  // state to record the timer on the right stream. 0 means no timer set
    +  private var rightTimerState: ValueState[Long] = _
    +
    +  private val leftRelativeSize: Long = -leftLowerBound
    +  private val rightRelativeSize: Long = leftUpperBound
    +
    +  private val relativeWindowSize = rightRelativeSize + leftRelativeSize
    +
    +  private var leftOperatorTime: Long = 0L
    +  private var rightOperatorTime: Long = 0L
    +
    +  private var backPressureSuggestion: Long = 0L
    +
    +  if (relativeWindowSize <= 0) {
    +    LOG.warn("The relative window size is non-positive, please check the 
join conditions.")
    +  }
    +
    +  if (allowedLateness < 0) {
    +    throw new IllegalArgumentException("The allowed lateness must be 
non-negative.")
    +  }
    +
    +
    +  /**
    +    * For holding back watermarks.
    +    *
    +    * @return the maximum delay for the outputs
    +    */
    +  def getMaxOutputDelay = Math.max(leftRelativeSize, rightRelativeSize) + 
allowedLateness;
    +
    +  /**
    +    * For dynamic query optimization.
    +    *
    +    * @return the suggested offset time for back-pressure
    +    */
    +  def getBackPressureSuggestion = backPressureSuggestion
    +
    +  override def open(config: Configuration) {
    +    val clazz = compile(
    +      getRuntimeContext.getUserCodeClassLoader,
    +      genJoinFuncName,
    +      genJoinFuncCode)
    +    joinFunction = clazz.newInstance()
    +
    +    cRowWrapper = new CRowWrappingCollector()
    +    cRowWrapper.setChange(true)
    +
    +    // Initialize the data caches.
    +    val leftListTypeInfo: TypeInformation[JList[Row]] = new 
ListTypeInfo[Row](leftType)
    +    val leftStateDescriptor: MapStateDescriptor[Long, JList[Row]] =
    +      new MapStateDescriptor[Long, JList[Row]](timeIndicator + 
"InnerJoinLeftCache",
    +        BasicTypeInfo.LONG_TYPE_INFO.asInstanceOf[TypeInformation[Long]], 
leftListTypeInfo)
    +    leftCache = getRuntimeContext.getMapState(leftStateDescriptor)
    +
    +    val rightListTypeInfo: TypeInformation[JList[Row]] = new 
ListTypeInfo[Row](rightType)
    +    val rightStateDescriptor: MapStateDescriptor[Long, JList[Row]] =
    +      new MapStateDescriptor[Long, JList[Row]](timeIndicator + 
"InnerJoinRightCache",
    +        BasicTypeInfo.LONG_TYPE_INFO.asInstanceOf[TypeInformation[Long]], 
rightListTypeInfo)
    +    rightCache = getRuntimeContext.getMapState(rightStateDescriptor)
    +
    +    // Initialize the timer states.
    +    val leftTimerStateDesc: ValueStateDescriptor[Long] =
    +      new ValueStateDescriptor[Long](timeIndicator + 
"InnerJoinLeftTimerState",
    +        classOf[Long])
    +    leftTimerState = getRuntimeContext.getState(leftTimerStateDesc)
    +
    +    val rightTimerStateDesc: ValueStateDescriptor[Long] =
    +      new ValueStateDescriptor[Long](timeIndicator + 
"InnerJoinRightTimerState",
    +        classOf[Long])
    +    rightTimerState = getRuntimeContext.getState(rightTimerStateDesc)
    +  }
    +
    +  /**
    +    * Process records from the left stream.
    +    *
    +    * @param cRowValue the input record
    +    * @param ctx       the context to register timer or get current time
    +    * @param out       the collector for outputting results
    +    *
    +    */
    +  override def processElement1(
    +    cRowValue: CRow,
    +    ctx: CoProcessFunction[CRow, CRow, CRow]#Context,
    +    out: Collector[CRow]): Unit = {
    +    val timeForRecord: Long = getTimeForRecord(ctx, cRowValue, true)
    +    getCurrentOperatorTime(ctx)
    +    processElement(
    +      cRowValue,
    +      timeForRecord,
    +      ctx,
    +      out,
    +      leftOperatorTime,
    +      rightOperatorTime,
    +      rightTimerState,
    +      leftCache,
    +      rightCache,
    +      true
    +    )
    +  }
    +
    +  /**
    +    * Process records from the right stream.
    +    *
    +    * @param cRowValue the input record
    +    * @param ctx       the context to get current time
    +    * @param out       the collector for outputting results
    +    *
    +    */
    +  override def processElement2(
    +    cRowValue: CRow,
    +    ctx: CoProcessFunction[CRow, CRow, CRow]#Context,
    +    out: Collector[CRow]): Unit = {
    +    val timeForRecord: Long = getTimeForRecord(ctx, cRowValue, false)
    +    getCurrentOperatorTime(ctx)
    +    processElement(
    +      cRowValue,
    +      timeForRecord,
    +      ctx,
    +      out,
    +      rightOperatorTime,
    +      leftOperatorTime,
    +      leftTimerState,
    +      rightCache,
    +      leftCache,
    +      false
    +    )
    +  }
    +
    +  /**
    +    * Put a record from the input stream into the cache and iterate the 
opposite cache to
    +    * output records meeting the join conditions. If there is no timer set 
for the OPPOSITE
    +    * STREAM, register one.
    +    */
    +  private def processElement(
    +    cRowValue: CRow,
    +    timeForRecord: Long,
    +    ctx: CoProcessFunction[CRow, CRow, CRow]#Context,
    +    out: Collector[CRow],
    +    myWatermark: Long,
    +    oppositeWatermark: Long,
    +    oppositeTimeState: ValueState[Long],
    +    recordListCache: MapState[Long, JList[Row]],
    +    oppositeCache: MapState[Long, JList[Row]],
    +    leftRecord: Boolean): Unit = {
    +    if (relativeWindowSize > 0) {
    +      //TODO Shall we consider adding a method for initialization with the 
context and collector?
    +      cRowWrapper.out = out
    +
    +      val record = cRowValue.row
    +
    +      //TODO Only if the time of the record is greater than the watermark, 
can we continue.
    +      if (timeForRecord >= myWatermark - allowedLateness) {
    +        val oppositeLowerBound: Long =
    +          if (leftRecord) timeForRecord - rightRelativeSize else 
timeForRecord - leftRelativeSize
    +
    +        val oppositeUpperBound: Long =
    +          if (leftRecord) timeForRecord + leftRelativeSize else 
timeForRecord + rightRelativeSize
    +
    +        // Put the record into the cache for later use.
    +        val recordList = if (recordListCache.contains(timeForRecord)) {
    +          recordListCache.get(timeForRecord)
    +        } else {
    +          new util.ArrayList[Row]()
    +        }
    +        recordList.add(record)
    +        recordListCache.put(timeForRecord, recordList)
    +
    +        // Register a timer on THE OTHER STREAM to remove records from the 
cache once they are
    +        // expired.
    +        if (oppositeTimeState.value == 0) {
    +          registerCleanUpTimer(
    +            ctx, timeForRecord, oppositeWatermark, oppositeTimeState, 
leftRecord, true)
    +        }
    +
    +        // Join the record with records from the opposite stream.
    +        val oppositeIterator = oppositeCache.iterator()
    +        var oppositeEntry: Entry[Long, util.List[Row]] = null
    +        var oppositeTime: Long = 0L;
    +        while (oppositeIterator.hasNext) {
    +          oppositeEntry = oppositeIterator.next
    +          oppositeTime = oppositeEntry.getKey
    +          if (oppositeTime < oppositeLowerBound - allowedLateness) {
    +            //TODO Considering the data out-of-order, we should not remove 
records here.
    +          } else if (oppositeTime >= oppositeLowerBound && oppositeTime <= 
oppositeUpperBound) {
    +            val oppositeRows = oppositeEntry.getValue
    +            var i = 0
    +            if (leftRecord) {
    +              while (i < oppositeRows.size) {
    +                joinFunction.join(record, oppositeRows.get(i), cRowWrapper)
    +                i += 1
    +              }
    +            } else {
    +              while (i < oppositeRows.size) {
    +                joinFunction.join(oppositeRows.get(i), record, cRowWrapper)
    +                i += 1
    +              }
    +            }
    +          } else if (oppositeTime > oppositeUpperBound) {
    +            //TODO If the keys are ordered, can we break here?
    +          }
    +        }
    +      } else {
    +        //TODO Need some extra logic here?
    +        LOG.warn(s"$record is out-of-date.")
    +      }
    +    }
    +  }
    +
    +  /**
    +    * Register a timer for cleaning up records in a specified time.
    +    *
    +    * @param ctx               the context to register timer
    +    * @param timeForRecord     time for the input record
    +    * @param oppositeWatermark watermark of the opposite stream
    +    * @param timerState        stores the timestamp for the next timer
    +    * @param leftRecord        record from the left or the right stream
    +    * @param firstTimer        whether this is the first timer
    +    */
    +  private def registerCleanUpTimer(
    +    ctx: CoProcessFunction[CRow, CRow, CRow]#Context,
    +    timeForRecord: Long,
    +    oppositeWatermark: Long,
    +    timerState: ValueState[Long],
    +    leftRecord: Boolean,
    +    firstTimer: Boolean): Unit = {
    +    val cleanUpTime = timeForRecord + (if (leftRecord) leftRelativeSize 
else rightRelativeSize) +
    +      allowedLateness + 1
    +    registerTimer(ctx, !leftRecord, cleanUpTime)
    +    LOG.debug(s"Register a clean up timer on the ${if (leftRecord) "RIGHT" 
else "LEFT"} state:"
    +      + s" timeForRecord = ${timeForRecord}, cleanUpTime = ${cleanUpTime}, 
oppositeWatermark = " +
    +      s"${oppositeWatermark}")
    +    timerState.update(cleanUpTime)
    +    if (cleanUpTime <= oppositeWatermark + allowedLateness && firstTimer) {
    +      backPressureSuggestion =
    +        if (leftRecord) (oppositeWatermark + allowedLateness - cleanUpTime)
    +        else -(oppositeWatermark + allowedLateness - cleanUpTime)
    +      LOG.warn("The clean timer for the " +
    +        s"${if (leftRecord) "left" else "right"}" +
    +        s" stream is lower than ${if (leftRecord) "right" else "left"} 
watermark." +
    +        s" requiredTime = ${formatTime(cleanUpTime)}, watermark = 
${formatTime(oppositeWatermark)},"
    +        + s"backPressureSuggestion = " + s"${backPressureSuggestion}.")
    +    }
    +  }
    +
    +
    +  /**
    +    * Called when a registered timer is fired.
    +    * Remove records which are earlier than the expiration time,
    +    * and register a new timer for the earliest remaining records.
    +    *
    +    * @param timestamp the timestamp of the timer
    +    * @param ctx       the context to register timer or get current time
    +    * @param out       the collector for returning result values
    +    */
    +  override def onTimer(
    +    timestamp: Long,
    +    ctx: CoProcessFunction[CRow, CRow, CRow]#OnTimerContext,
    +    out: Collector[CRow]): Unit = {
    +    getCurrentOperatorTime(ctx)
    +    //TODO In the future, we should separate the left and right 
watermarks. Otherwise, the
    +    //TODO registered timer of the faster stream will be delayed, even if 
the watermarks have
    +    //TODO already been emitted by the source.
    +    if (leftTimerState.value == timestamp) {
    +      val rightExpirationTime = leftOperatorTime - rightRelativeSize - 
allowedLateness - 1
    +      removeExpiredRecords(
    +        timestamp,
    +        rightExpirationTime,
    +        leftOperatorTime,
    +        rightCache,
    +        leftTimerState,
    +        ctx,
    +        false
    +      )
    +    }
    +
    +    if (rightTimerState.value == timestamp) {
    +      val leftExpirationTime = rightOperatorTime - leftRelativeSize - 
allowedLateness - 1
    +      removeExpiredRecords(
    +        timestamp,
    +        leftExpirationTime,
    +        rightOperatorTime,
    +        leftCache,
    +        rightTimerState,
    +        ctx,
    +        true
    +      )
    +    }
    +  }
    +
    +  /**
    +    * Remove the expired records. Register a new timer if the cache still 
holds records
    +    * after the cleaning up.
    +    */
    +  private def removeExpiredRecords(
    +    timerFiringTime: Long,
    +    expirationTime: Long,
    +    oppositeWatermark: Long,
    +    recordCache: MapState[Long, JList[Row]],
    +    timerState: ValueState[Long],
    +    ctx: CoProcessFunction[CRow, CRow, CRow]#OnTimerContext,
    +    removeLeft: Boolean): Unit = {
    +
    +    val keysIterator = recordCache.keys().iterator()
    +
    +    // Search for expired timestamps.
    +    // If we find a non-expired timestamp, remember the timestamp and 
leave the loop.
    +    // This way we find all expired timestamps if they are sorted without 
doing a full pass.
    +    var earliestTimestamp: Long = -1L
    +    var recordTime: Long = 0L
    +    while (keysIterator.hasNext) {
    +      //TODO The "short-circuit" code was commented, because when using a 
StateMap with
    +      //TODO unordered keys, the cache will grow indefinitely!
    +      // && earliestTimestamp < 0) {
    +      recordTime = keysIterator.next
    +      if (recordTime <= expirationTime) {
    +        // TODO Not sure if we can remove records directly.
    +        keysIterator.remove()
    +      } else {
    +        // We find the earliest timestamp that is still valid.
    +        if (recordTime < earliestTimestamp || earliestTimestamp < 0) {
    +          earliestTimestamp = recordTime
    +        }
    +      }
    +    }
    +    // If the cache contains non-expired timestamps, register a new timer.
    +    // Otherwise clear the states.
    +    if (earliestTimestamp > 0) {
    +      registerCleanUpTimer(ctx, earliestTimestamp, oppositeWatermark, 
timerState, removeLeft, false)
    --- End diff --
    
    That's indeed a critical problem. The current approach is designed for 
minimizing the cached records, but may spend too much time in cleaning up. I 
think of two ways to improve that: (1) as you suggested, we delay the timers 
with specific spans, or (2) we schedule the cleaning up process with a specific 
interval according to the relative window size. What do you think?
    Actually, every problem will be solved if we got a state with ordered keys, 
like the TreeMap ;-)


---

Reply via email to