[ https://issues.apache.org/jira/browse/FLINK-9715?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16651837#comment-16651837 ]
ASF GitHub Bot commented on FLINK-9715: --------------------------------------- pnowojski 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_r225570366 ########## File path: flink-libraries/flink-table/src/main/scala/org/apache/flink/table/runtime/join/TemporalRowtimeJoin.scala ########## @@ -0,0 +1,339 @@ +/* + * 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 +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 org.apache.flink.util.Preconditions.checkState + +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`). + */ Review comment: @kl0u other thing is that I don't fully understand our pressure on using RocksDB, while vast majority of our users (at least that was result of my analysis of `DataStream` usage) would be perfectly fine with Heap state backend. And the general trend in big data (or any large scale systems) is trying to do as much as you can in memory. Regarding performance penalty, just accessing one more level of collection is a significant slowdown (more memory access), even assuming that we could reuse the Lists - which we couldn't, since keeping the empty list on the state would prevent state cleanup. Regarding the readability, I agree, however I wouldn't consider `Map<Timestamp, List<CRow>>` as more readable. It has similar "mental" complexity. The structure that we need here is `OrderedMultiSet` or `OrderedMultiMap` (ordered by timestamp), which would be significantly more efficient and more natural/easier to understand, without the need for binary search etc. ---------------------------------------------------------------- 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)