大家都知道在spark1.3版本后,kafkautil里面提供了两个创建dstream的方法,一个是老版本中有的createStream方法,还有一个是后面新加的createDirectStream方法。关于这两个方法的优缺点,官方已经说的很详细(http:///docs/latest/streaming-kafka-integration.html),总之就是createDirectStream性能会更好一点,通过新方法创建出来的dstream的rdd partition和kafka的topic的partition是一一对应的,通过低阶API直接从kafka的topic消费消息,但是它不再往zookeeper中更新consumer offsets,使得基于zk的consumer offsets的监控工具都会失效。
官方只是蜻蜓点水般的说了一下可以在foreachRDD中更新zookeeper上的offsets:
[plain] view plain copy
1. directKafkaStream.foreachRDD { rdd =>
2. val offsetRanges = rdd.asInstanceOf[HasOffsetRanges]
3. // offsetRanges.length = # of Kafka partitions being consumed
4. ...
5. }对应
Exactly-once semantics要自己去实现了,大致的实现思路就是在driver启动的时候先从zk上获得consumer offsets信息,createDirectStream有两个重载方法,其中一个可以设置从任意offsets位置开始消费,部分代码如下:
[plain] view plain copy
1. def createDirectStream(implicit streamingConfig: StreamingConfig, kc: KafkaCluster) = {
2.
3. val extractors = streamingConfig.getExtractors()
4. //从zookeeper上读取offset开始消费message
5. val messages = {
6. val kafkaPartitionsE = kc.getPartitions(streamingConfig.topicSet)
7. if (kafkaPartitionsE.isLeft) throw new SparkException("get kafka partition failed:")
8. val kafkaPartitions = kafkaPartitionsE.right.get
9. val consumerOffsetsE = kc.getConsumerOffsets(streamingConfig.group, kafkaPartitions)
10. if (consumerOffsetsE.isLeft) throw new SparkException("get kafka consumer offsets failed:")
11. val consumerOffsets = consumerOffsetsE.right.get
12. consumerOffsets.foreach {
13. case (tp, n) => println("===================================" + tp.topic + "," + tp.partition + "," + n)
14. }
15. KafkaUtils.createDirectStream[String, String, StringDecoder, StringDecoder, (String, String)](
16. ssc, kafkaParams, consumerOffsets, (mmd: MessageAndMetadata[String, String]) => (mmd.key, mmd.message))
17. }
18. messages
19. }
这里会有几个问题,就是在一个group是新的consumer group时,即首次消费,zk上海没有相应的group offsets目录,这时要先初始化一下zk上的offsets目录,或者是zk上记录的offsets已经过时,由于kafka有定时清理策略,直接从zk上的offsets开始消费会报ArrayOutofRange异常,即找不到offsets所属的index文件了,针对这两种情况,做了以下处理:
[plain] view plain copy
1. def setOrUpdateOffsets(implicit streamingConfig: StreamingConfig, kc: KafkaCluster): Unit = {
2. streamingConfig.topicSet.foreach(topic => {
3. println("current topic:" + topic)
4. var hasConsumed = true
5. val kafkaPartitionsE = kc.getPartitions(Set(topic))
6. if (kafkaPartitionsE.isLeft) throw new SparkException("get kafka partition failed:")
7. val kafkaPartitions = kafkaPartitionsE.right.get
8. val consumerOffsetsE = kc.getConsumerOffsets(streamingConfig.group, kafkaPartitions)
9. if (consumerOffsetsE.isLeft) hasConsumed = false
10. if (hasConsumed) {
11. //如果有消费过,有两种可能,如果streaming程序执行的时候出现kafka.common.OffsetOutOfRangeException,说明zk上保存的offsets已经过时了,即kafka的定时清理策略已经将包含该offsets的文件删除。
12. //针对这种情况,只要判断一下zk上的consumerOffsets和leaderEarliestOffsets的大小,如果consumerOffsets比leaderEarliestOffsets还小的话,说明是过时的offsets,这时把leaderEarliestOffsets更新为consumerOffsets
13. val leaderEarliestOffsets = kc.getEarliestLeaderOffsets(kafkaPartitions).right.get
14. println(leaderEarliestOffsets)
15. val consumerOffsets = consumerOffsetsE.right.get
16. val flag = consumerOffsets.forall {
17. case (tp, n) => n < leaderEarliestOffsets(tp).offset
18. }
19. if (flag) {
20. println("consumer group:" + streamingConfig.group + " offsets已经过时,更新为leaderEarliestOffsets")
21. val offsets = leaderEarliestOffsets.map {
22. case (tp, offset) => (tp, offset.offset)
23. }
24. kc.setConsumerOffsets(streamingConfig.group, offsets)
25. }
26. else {
27. println("consumer group:" + streamingConfig.group + " offsets正常,无需更新")
28. }
29. }
30. else {
31. //如果没有被消费过,则从最新的offset开始消费。
32. val leaderLatestOffsets = kc.getLatestLeaderOffsets(kafkaPartitions).right.get
33. println(leaderLatestOffsets)
34. println("consumer group:" + streamingConfig.group + " 还未消费过,更新为leaderLatestOffsets")
35. val offsets = leaderLatestOffsets.map {
36. case (tp, offset) => (tp, offset.offset)
37. }
38. kc.setConsumerOffsets(streamingConfig.group, offsets)
39. }
40. })
41. }
这里又碰到了一个问题,从consumer offsets到leader latest offsets中间延迟了很多消息,在下一次启动的时候,首个batch要处理大量的消息,会导致spark-submit设置的资源无法满足大量消息的处理而导致崩溃。因此在spark-submit启动的时候多加了一个配置:--conf spark.streaming.kafka.maxRatePerPartition=10000。限制每秒钟从topic的每个partition最多消费的消息条数,这样就把首个batch的大量的消息拆分到多个batch中去了,为了更快的消化掉delay的消息,可以调大计算资源和把这个参数调大。
OK,driver启动的问题解决了,那么接下来处理处理完消息后更新zk offsets的工作,这里要注意是在处理完之后再更新,想想如果你消费了消息先更新zk offset在去处理消息将处理好的消息保存到其他地方去,如果后一步由于处理消息的代码有BUG失败了,前一步已经更新了zk了,会导致这部分消息虽然被消费了但是没被处理,等你把处理消息的BUG修复再重新提交后,这部分消息在下次启动的时候不会再被消费了,因为你已经更新了ZK OFFSETS,针对这些因素考虑,部分代码实现如下:
[plain] view plain copy
仔细想一想,还是没有实现精确一次的语义,写入mongo和更新ZK由于不是一个事务的,如果更新mongo成功,然后更新ZK失败,则下次启动的时候这个批次的数据就被重复计算,对于UV由于是addToSet去重操作,没什么影响,但是PV是inc操作就会多算这一个批次的的数据,其实如果batch time比较短的话,其实都还是可以接受的。
1. def updateZKOffsets(rdd: RDD[(String, String)])(implicit streamingConfig: StreamingConfig, kc: KafkaCluster): Unit = {
2. println("rdd not empty,update zk offset")
3. val offsetsList = rdd.asInstanceOf[HasOffsetRanges].offsetRanges
4.
5.
6. for (offsets <- offsetsList) {
7. val topicAndPartition = TopicAndPartition(offsets.topic, offsets.partition)
8. val o = kc.setConsumerOffsets(streamingConfig.group, Map((topicAndPartition, offsets.untilOffset)))
9. if (o.isLeft) {
10. println(s"Error updating the offset to Kafka cluster: ${o.left.get}")
11. }
12. }
13. }
14.
15. def processData(messages: InputDStream[(String, String)])(implicit streamingConfig: StreamingConfig, kc: KafkaCluster): Unit = {
16. messages.foreachRDD(rdd => {
17. if (!rdd.isEmpty()) {
18.
19. val datamodelRDD = streamingConfig.relation match {
20. case "1" =>
21. val (topic, _) = streamingConfig.topic_table_mapping
22. val extractor = streamingConfig.getExtractor(topic)
23. // Create direct kafka stream with brokers and topics
24. val topicsSet = Set(topic)
25. val datamodel = rdd.filter(msg => {
26. extractor.filter(msg)
27. }).map(msg => extractor.msgToRow(msg))
28. datamodel
29. case "2" =>
30. val (topics, _) = streamingConfig.topic_table_mapping
31. val extractors = streamingConfig.getExtractors(topics)
32. val topicsSet = topics.split(",").toSet
33.
34. //kafka msg为key-value形式,key用来对msg进行分区用的,为了散列存储消息,采集器那边key采用的是:topic|加一个随机数的形式,例如:rd_e_pal|20,split by |取0可以拿到对应的topic名字,这样union在一起的消息可以区分出来自哪一个topic
35. val datamodel = rdd.filter(msg => {
36. //kafka msg为key-value形式,key用来对msg进行分区用的,为了散列存储消息,采集器那边key采用的是:topic|加一个随机数的形式,例如:rd_e_pal|20,split by |取0可以拿到对应的topic名字,这样union在一起的消息可以区分出来自哪一个topic
37. val keyValid = msg != null && msg._1 != null && msg._1.split("\\|").length == 2
38. if (keyValid) {
39. val topic = msg._1.split("\\|")(0)
40. val (_, extractor) = extractors.find(p => {
41. p._1.equalsIgnoreCase(topic)
42. }).getOrElse(throw new RuntimeException("配置文件中没有找到topic:" + topic + " 对应的extractor"))
43. //trim去掉末尾的换行符,否则取最后一个字段时会有一个\n
44. extractor.filter(msg._2.trim)
45. }
46. else {
47. false
48. }
49.
50. }).map {
51. case (key, msgContent) =>
52. val topic = key.split("\\|")(0)
53. val (_, extractor) = extractors.find(p => {
54. p._1.equalsIgnoreCase(topic)
55. }).getOrElse(throw new RuntimeException("配置文件中没有找到topic:" + topic + " 对应的extractor"))
56. extractor.msgToRow((key, msgContent))
57. }
58. datamodel
59. }
60. //先处理消息
61. processRDD(datamodelRDD)
62. //再更新offsets
63. updateZKOffsets(rdd)
64. }
65. })
66. }
67.
68. def processRDD(rdd: RDD[Row])(implicit streamingConfig: StreamingConfig) = {
69. if (streamingConfig.targetType == "mongo") {
70. val target = streamingConfig.getTarget().asInstanceOf[MongoTarget]
71. if (!MongoDBClient.db.collectionExists(target.collection)) {
72. println("create collection:" + target.collection)
73. MongoDBClient.db.createCollection(target.collection, MongoDBObject("storageEngine" -> MongoDBObject("wiredTiger" -> MongoDBObject())))
74. val coll = MongoDBClient.db(target.collection)
75. //创建ttl index
76. if (target.ttlIndex) {
77. val indexs = coll.getIndexInfo
78. if (indexs.find(p => p.get("name") == "ttlIndex") == None) {
79. coll.createIndex(MongoDBObject(target.ttlColumn -> 1), MongoDBObject("expireAfterSeconds" -> target.ttlExpire, "name" -> "ttlIndex"))
80. }
81. }
82. }
83.
84. }
85.
86. val (_, table) = streamingConfig.topic_table_mapping
87. val schema = streamingConfig.getTableSchema(table)
88.
89. // Get the singleton instance of SQLContext
90. val sqlContext = HIVEContextSingleton.getInstance(rdd.sparkContext)
91.
92. // Convert RDD[String] to RDD[case class] to DataFrame
93. val dataFrame = sqlContext.createDataFrame(rdd, schema)
94.
95. // Register as table
96. dataFrame.registerTempTable(table)
97.
98. // Do word count on table using SQL and print it
99. val results = sqlContext.sql(streamingConfig.sql)
100. //select dt,hh(vtm) as hr,app_key, collect_set(device_id) as deviceids from rd_e_app_header where dt=20150401 and hh(vtm)='01' group by dt,hh(vtm),app_key limit 100 ;
101. // results.show()
102. streamingConfig.targetType match {
103. case "mongo" => saveToMongo(results)
104. case "show" => results.show()
105. }
106.
107. }
108.
109.
110. def saveToMongo(df: DataFrame)(implicit streamingConfig: StreamingConfig) = {
111. val target = streamingConfig.getTarget().asInstanceOf[MongoTarget]
112. val coll = MongoDBClient.db(target.collection)
113. val result = df.collect()
114. if (result.size > 0) {
115. val bulkWrite = coll.initializeUnorderedBulkOperation
116. result.foreach(row => {
117. val id = row(target.pkIndex)
118. val setFields = target.columns.filter(p => p.op == "set").map(f => (, row(f.index))).toArray
119. val incFields = target.columns.filter(p => p.op == "inc").map(f => {
120. (, row(f.index).asInstanceOf[Long])
121. }).toArray
122. // obj=obj.++($addToSet(MongoDBObject("test"->MongoDBObject("$each"->Array(3,4)),"test1"->MongoDBObject("$each"->Array(1,2)))))
123. var obj = MongoDBObject()
124. var addToSetObj = MongoDBObject()
125. target.columns.filter(p => p.op == "addToSet").foreach(col => {
126. col.mType match {
127. case "Int" =>
128. addToSetObj = addToSetObj.++( -> MongoDBObject("$each" -> row(col.index).asInstanceOf[ArrayBuffer[Int]]))
129. case "Long" =>
130. addToSetObj = addToSetObj.++( -> MongoDBObject("$each" -> row(col.index).asInstanceOf[ArrayBuffer[Long]]))
131. case "String" =>
132. addToSetObj = addToSetObj.++( -> MongoDBObject("$each" -> row(col.index).asInstanceOf[ArrayBuffer[String]]))
133. }
134.
135. })
136. if (addToSetObj.size > 0) obj = obj.++($addToSet(addToSetObj))
137. if (incFields.size > 0) obj = obj.++($inc(incFields: _*))
138. if (setFields.size > 0) obj = obj.++($set(setFields: _*))
139. bulkWrite.find(MongoDBObject("_id" -> id)).upsert().updateOne(obj)
140. })
141. bulkWrite.execute()
142. }
143. }
















