[
https://issues.apache.org/jira/browse/FLINK-9715?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16646603#comment-16646603
]
ASF GitHub Bot commented on FLINK-9715:
---------------------------------------
hequn8128 commented on a change in pull request #6776: [FLINK-9715][table]
Support temporal join with event time
URL: https://github.com/apache/flink/pull/6776#discussion_r224492754
##########
File path:
flink-libraries/flink-table/src/main/scala/org/apache/flink/table/runtime/join/TemporalRowtimeJoin.scala
##########
@@ -0,0 +1,326 @@
+/*
+ * 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.lang.{Long => JLong}
+import java.util
+import java.util.Comparator
+
+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.runtime.state.{VoidNamespace, VoidNamespaceSerializer}
+import org.apache.flink.streaming.api.SimpleTimerService
+import org.apache.flink.streaming.api.operators._
+import org.apache.flink.streaming.runtime.streamrecord.StreamRecord
+import org.apache.flink.table.api.{StreamQueryConfig, TableException}
+import org.apache.flink.table.codegen.Compiler
+import org.apache.flink.table.runtime.CRowWrappingCollector
+import org.apache.flink.table.runtime.types.CRow
+import org.apache.flink.table.typeutils.TypeCheckUtils._
+import org.apache.flink.table.util.Logging
+import org.apache.flink.types.Row
+
+import scala.collection.JavaConversions._
+
+/**
+ * This operator works by keeping on the state collection of probe and build
records to process
+ * on next watermark. The idea is that between watermarks we are collecting
those elements
+ * and once we are sure that there will be no updates we emit the correct
result and clean up the
+ * state.
+ *
+ * Cleaning up the state drops all of the "old" values from the probe side,
where "old" is defined
+ * as older then the current watermark. Build side is also cleaned up in the
similar fashion,
+ * however we always keep at least one record - the latest one - even if it's
past the last
+ * watermark.
+ *
+ * One more trick is how the emitting results and cleaning up is triggered.
It is achieved
+ * by registering timers for the keys. We could register a timer for every
probe and build
+ * side element's event time (when watermark exceeds this timer, that's when
we are emitting and/or
+ * cleaning up the state). However this would cause huge number of registered
timers. For example
+ * with following evenTimes of probe records accumulated: {1, 2, 5, 8, 9}, if
we
+ * had received Watermark(10), it would trigger 5 separate timers for the
same key. To avoid that
+ * we always keep only one single registered timer for any given key,
registered for the minimal
+ * value. Upon triggering it, we process all records with event times older
then or equal to
+ * currentWatermark.
+ */
+class TemporalRowtimeJoin(
+ leftType: TypeInformation[Row],
+ rightType: TypeInformation[Row],
+ genJoinFuncName: String,
+ genJoinFuncCode: String,
+ queryConfig: StreamQueryConfig,
+ leftTimeAttribute: Int,
+ rightTimeAttribute: Int)
+ extends AbstractStreamOperator[CRow]
+ with TwoInputStreamOperator[CRow, CRow, CRow]
+ with Triggerable[Any, VoidNamespace]
+ with Compiler[FlatJoinFunction[Row, Row, Row]]
+ with Logging {
+
+ validateEqualsHashCode("join", leftType)
+ validateEqualsHashCode("join", rightType)
+
+ private val NEXT_LEFT_INDEX_STATE_NAME = "next-index"
+ private val LEFT_STATE_NAME = "left"
+ private val RIGHT_STATE_NAME = "right"
+ private val REGISTERED_TIMER_STATE_NAME = "timer"
+ private val TIMERS_STATE_NAME = "timers"
+
+ private val rightRowtimeComparator = new
RowtimeComparator(rightTimeAttribute)
+
+ /**
+ * Incremental index generator for `leftState`'s keys.
+ */
+ private var nextLeftIndex: ValueState[JLong] = _
+
+ /**
+ * Mapping from artificial row index (generated by `nextLeftIndex`) into
the left side `Row`.
+ * We can not use List to accumulate Rows, because we need efficient
deletes of the oldest rows.
+ *
+ * TODO: this could be OrderedMultiMap[Jlong, Row] indexed by row's
timestamp, to avoid
+ * full map traversals (if we have lots of rows on the state that exceed
`currentWatermark`).
+ */
+ private var leftState: MapState[JLong, Row] = _
+
+ /**
+ * Mapping from timestamp to right side `Row`.
+ *
+ * TODO: having `rightState` as an OrderedMapState would allow us to avoid
sorting cost
+ * once per watermark
+ */
+ private var rightState: MapState[JLong, Row] = _
+
+ private var registeredTimer: ValueState[JLong] = _ // JLong for correct
handling of default null
+
+ private var cRowWrapper: CRowWrappingCollector = _
+ private var collector: TimestampedCollector[CRow] = _
+ private var timerService: SimpleTimerService = _
+
+ private var joinFunction: FlatJoinFunction[Row, Row, Row] = _
+
+ override def open(): Unit = {
+ val clazz = compile(
+ getRuntimeContext.getUserCodeClassLoader,
+ genJoinFuncName,
+ genJoinFuncCode)
+
+ joinFunction = clazz.newInstance()
+
+ nextLeftIndex = getRuntimeContext.getState(
+ new ValueStateDescriptor[JLong](NEXT_LEFT_INDEX_STATE_NAME,
BasicTypeInfo.LONG_TYPE_INFO))
+ leftState = getRuntimeContext.getMapState(
+ new MapStateDescriptor[JLong, Row](LEFT_STATE_NAME,
BasicTypeInfo.LONG_TYPE_INFO, leftType))
+ rightState = getRuntimeContext.getMapState(
+ new MapStateDescriptor[JLong, Row](RIGHT_STATE_NAME,
BasicTypeInfo.LONG_TYPE_INFO, rightType))
+ registeredTimer = getRuntimeContext.getState(
+ new ValueStateDescriptor[JLong](REGISTERED_TIMER_STATE_NAME,
BasicTypeInfo.LONG_TYPE_INFO))
+
+ collector = new TimestampedCollector[CRow](output)
+ cRowWrapper = new CRowWrappingCollector()
+ cRowWrapper.out = collector
+ cRowWrapper.setChange(true)
+
+ val internalTimerService = getInternalTimerService(
+ TIMERS_STATE_NAME,
+ VoidNamespaceSerializer.INSTANCE,
+ this)
+
+ timerService = new SimpleTimerService(internalTimerService)
+ }
+
+ override def processElement1(element: StreamRecord[CRow]): Unit = {
+ if (!element.getValue.change) {
+ throw new TableException(
+ s"${classOf[TemporalRowtimeJoin].getSimpleName} does not support
retractions on the " +
+ s"left side.")
+ }
+
+ leftState.put(getNextLeftIndex, element.getValue.row)
+ registerSmallestTimer(getLeftTime(element.getValue.row)) // Timer to emit
and clean up the state
+ }
+
+ override def processElement2(element: StreamRecord[CRow]): Unit = {
+ if (!element.getValue.change) {
+ throw new TableException(
Review comment:
RowTime UpsertSource may make it possible. It would be good to validate
during `translateToPlan` before running on a cluster.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
> Support versioned joins with event time
> ---------------------------------------
>
> Key: FLINK-9715
> URL: https://issues.apache.org/jira/browse/FLINK-9715
> Project: Flink
> Issue Type: Sub-task
> Components: Table API & SQL
> Affects Versions: 1.5.0
> Reporter: Piotr Nowojski
> Assignee: Piotr Nowojski
> Priority: Major
> Labels: pull-request-available
>
> Queries like:
> {code:java}
> SELECT
> o.amount * r.rate
> FROM
> Orders AS o,
> LATERAL TABLE (Rates(o.rowtime)) AS r
> WHERE o.currency = r.currency{code}
> should work with event time
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
