Flink中的多source+event watermark测试
这次需要做一个监控项目，全网日志的指标计算，上线的话，计算量应该是百亿/天 单个source对应的sql如下 最原始的sql select pro,throwable,level,ip,`count`,id,`time`,firstl,lastl from ( select pro,throwable,level,ip, count(*) as `count`, lastStrInGroupSkipNull(CONCAT_WS('_',KAFKA_TOPIC,CAST(KAFKA_PARTITION AS VARCHAR),CAST(KAFKA_OFFSET as VARCHAR))) as id, firstLong(l) as firstl, lastLong(l) as lastl, TUMBLE_END(SPT, INTERVAL '3' SECOND) as `time` from input.`ymm-appmetric-dev-self1` where pro IS NOT NULL and throwable IS NOT NULL and level IS NOT NULL and level='ERROR' and ip IS NOT NULL group by pro,throwable,level,ip,TUMBLE(SPT,INTERVAL '3' SECOND) ) where 1=uniqueWithin100MS(pro,throwable,level,ip,`time`) ---先做技术论证，写了下面一个sql select pro,throwable,level,ip,`count`,id,`time`,firstl,lastl from ( select pro,throwable,level,ip,count(*) as `count`, lastStrInGroupSkipNull(CONCAT_WS('_',KAFKA_TOPIC,CAST(KAFKA_PARTITION AS VARCHAR),CAST(KAFKA_OFFSET as VARCHAR))) as id, firstLong(l) as firstl, lastLong(l) as lastl, TUMBLE_END(SPT, INTERVAL '3' SECOND) as `time` from ( select pro,throwable,level,ip from input.`ymm-appmetric-dev-self1` where pro IS NOT NULL and throwable IS NOT NULL and level IS NOT NULL and level='ERROR' and ip IS NOT NULL union select pro,throwable,level,ip from input.`ymm-appmetric-dev-self2` where pro IS NOT NULL and throwable IS NOT NULL and level IS NOT NULL and level='ERROR' and ip IS NOT NULL ) group by pro,throwable,level,ip,TUMBLE(SPT,INTERVAL '3' SECOND) ) where 1=uniqueWithin100MS(pro,throwable,level,ip,`time`) 然后拉起flink任务,观察是否可顺利启动---果然报错了 Caused by: org.apache.calcite.sql.validate.SqlValidatorException: Column 'SPT' not found in any table 定位一下，看看是什么问题导致的,看了下之前写的sql，猜测是因为UNION的时候，没有在每个表里带上SPT时间属性字段以及其它字段，补上后sql如下 select pro,throwable,level,ip,`count`,id,`time`,firstl,lastl from ( select pro,throwable,level,ip,count(*) as `count`, lastStrInGroupSkipNull(CONCAT_WS('_',KAFKA_TOPIC,CAST(KAFKA_PARTITION AS VARCHAR),CAST(KAFKA_OFFSET as VARCHAR))) as id, firstLong(l) as firstl, lastLong(l) as lastl, TUMBLE_END(SPT, INTERVAL '3' SECOND) as `time` from ( select pro,throwable,level,ip,l,KAFKA_TOPIC,KAFKA_PARTITION,KAFKA_OFFSET,SPT from input.`ymm-appmetric-dev-self1` where pro IS NOT NULL and throwable IS NOT NULL and level IS NOT NULL and level='ERROR' and ip IS NOT NULL union select pro,throwable,level,ip,l,KAFKA_TOPIC,KAFKA_PARTITION,KAFKA_OFFSET,SPT from input.`ymm-appmetric-dev-self2` where pro IS NOT NULL and throwable IS NOT NULL and level IS NOT NULL and level='ERROR' and ip IS NOT NULL ) group by pro,throwable,level,ip,TUMBLE(SPT,INTERVAL '3' SECOND) ) where 1=uniqueWithin100MS(pro,throwable,level,ip,`time`) 再重启看看，这次应该差不多了吧---sql可以顺利编译，但是还是有错 奇怪了，之前并没有这样的错误，赞，我们来看看问题在哪！ 我们打开类的层次图如下 借这个机会加强对这些类的理解！ ---经过我的调试，发现问题出现在union上，不加这个Union，啥事没有；加了就报错，下面我们再回到调用栈看看 一个人调试了一个下午，-_-||，最终发现知道修改一个地方就行 union -> union all 厉害了，给大佬低头！ ----好，既然解决了，我们继续来debug原理层！ 测试了一下，发现多source跟单source相比，单source的watermark很好理解，但是多source就稍微复杂些，下面我们来研究下原理！ 首先，观察一下现有的图，如下所示： 下面再来研究一下线程，jstack一把 我们来分析上面的线程，看看有没有收获！挑几个重点线程讲解 "VM Periodic Task Thread" os_prio=0 tid=0x00007f366825e800 nid=0x63d waiting on condition 百度可以知道 该线程是JVM周期性任务调度的线程，它由WatcherThread创建，是一个单例对象。该线程在JVM内使用得比较频繁，比如：定期的内存监控、JVM运行状况监控。 下面几个是GC线程 "Gang worker#0 (Parallel GC Threads)" os_prio=0 tid=0x00007f3668031800 nid=0x626 runnable "Gang worker#1 (Parallel GC Threads)" os_prio=0 tid=0x00007f3668033800 nid=0x627 runnable "Gang worker#2 (Parallel GC Threads)" os_prio=0 tid=0x00007f3668035800 nid=0x628 runnable "Gang worker#3 (Parallel GC Threads)" os_prio=0 tid=0x00007f3668037800 nid=0x629 runnable "Gang worker#4 (Parallel GC Threads)" os_prio=0 tid=0x00007f3668039800 nid=0x62a runnable "Gang worker#5 (Parallel GC Threads)" os_prio=0 tid=0x00007f366803b000 nid=0x62b runnable "Gang worker#6 (Parallel GC Threads)" os_prio=0 tid=0x00007f366803d000 nid=0x62c runnable "Gang worker#7 (Parallel GC Threads)" os_prio=0 tid=0x00007f366803f000 nid=0x62d runnable "Concurrent Mark-Sweep GC Thread" os_prio=0 tid=0x00007f36680b7000 nid=0x630 runnable "Gang worker#0 (Parallel CMS Threads)" os_prio=0 tid=0x00007f36680b2800 nid=0x62e runnable "Gang worker#1 (Parallel CMS Threads)" os_prio=0 tid=0x00007f36680b4800 nid=0x62f runnable --- "main" #1 prio=5 os_prio=0 tid=0x00007f3668019800 nid=0x625 waiting on condition [0x00007f3670010000] 主线程，在flink内部等待所有事情结束 "New I/O worker #1" #24 prio=5 os_prio=0 tid=0x00007f366995f000 nid=0x648 runnable [0x00007f3642cd1000] 内部netty线程 --- "Source: MyKafka010JsonTableSource -> from: (l, KAFKA_TOPIC, KAFKA_PARTITION, KAFKA_OFFSET, pro, throwable, level, ip, SPT) -> Timestamps/Watermarks -> where: (AND(=(level, _UTF-16LE'ERROR'), IS NOT NULL(pro), IS NOT NULL(throwable), IS NOT NULL(ip))), select: (pro, throwable, CAST(_UTF-16LE'ERROR') AS level, ip, SPT, CONCAT_WS(_UTF-16LE'_', KAFKA_TOPIC, CAST(KAFKA_PARTITION), CAST(KAFKA_OFFSET)) AS $f5, l) (1/1)" #51 prio=5 os_prio=0 tid=0x00007f363d11a800 nid=0x65e in Object.wait() [0x00007f3641ac3000] java.lang.Thread.State: WAITING (on object monitor) at java.lang.Object.wait(Native Method) at java.lang.Object.wait(Object.java:502) at org.apache.flink.streaming.connectors.kafka.internal.Handover.pollNext(Handover.java:74) - locked <0x00000000e6ee2df0> (a java.lang.Object) at org.apache.flink.streaming.connectors.kafka.internal.Kafka09Fetcher.runFetchLoop(Kafka09Fetcher.java:133) at org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumerBase.run(FlinkKafkaConsumerBase.java:721) at org.apache.flink.streaming.api.operators.StreamSource.run(StreamSource.java:87) at org.apache.flink.streaming.api.operators.StreamSource.run(StreamSource.java:56) at org.apache.flink.streaming.runtime.tasks.SourceStreamTask.run(SourceStreamTask.java:99) at org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:306) at org.apache.flink.runtime.taskmanager.Task.run(Task.java:703) at java.lang.Thread.run(Thread.java:748) "Source: MyKafka010JsonTableSource -> from: (l, KAFKA_TOPIC, KAFKA_PARTITION, KAFKA_OFFSET, pro, throwable, level, ip, SPT) -> Timestamps/Watermarks -> where: (AND(=(level, _UTF-16LE'ERROR'), IS NOT NULL(pro), IS NOT NULL(throwable), IS NOT NULL(ip))), select: (pro, throwable, CAST(_UTF-16LE'ERROR') AS level, ip, SPT, CONCAT_WS(_UTF-16LE'_', KAFKA_TOPIC, CAST(KAFKA_PARTITION), CAST(KAFKA_OFFSET)) AS $f5, l) (1/1)" #50 prio=5 os_prio=0 tid=0x00007f363d120800 nid=0x65d in Object.wait() [0x00007f3641bc4000] java.lang.Thread.State: WAITING (on object monitor) at java.lang.Object.wait(Native Method) at java.lang.Object.wait(Object.java:502) at org.apache.flink.streaming.connectors.kafka.internal.Handover.pollNext(Handover.java:74) - locked <0x00000000e6ee2e98> (a java.lang.Object) at org.apache.flink.streaming.connectors.kafka.internal.Kafka09Fetcher.runFetchLoop(Kafka09Fetcher.java:133) at org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumerBase.run(FlinkKafkaConsumerBase.java:721) at org.apache.flink.streaming.api.operators.StreamSource.run(StreamSource.java:87) at org.apache.flink.streaming.api.operators.StreamSource.run(StreamSource.java:56) at org.apache.flink.streaming.runtime.tasks.SourceStreamTask.run(SourceStreamTask.java:99) at org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:306) at org.apache.flink.runtime.taskmanager.Task.run(Task.java:703) at java.lang.Thread.run(Thread.java:748) 有2个线程是用来获取消息，对于这2个线程来说，这2个消息不是直接读取kafka,而是其它线程读取kafka喂给这2个线程 --- "time attribute: (SPT) (1/1)" #53 prio=5 os_prio=0 tid=0x00007f363d8e4000 nid=0x662 in Object.wait() [0x00007f36418c1000] java.lang.Thread.State: WAITING (on object monitor) at java.lang.Object.wait(Native Method) at java.lang.Object.wait(Object.java:502) at org.apache.flink.runtime.io.network.partition.consumer.UnionInputGate.waitAndGetNextInputGate(UnionInputGate.java:205) - locked <0x00000000e6ee8210> (a java.util.ArrayDeque) at org.apache.flink.runtime.io.network.partition.consumer.UnionInputGate.getNextBufferOrEvent(UnionInputGate.java:163) at org.apache.flink.streaming.runtime.io.BarrierTracker.getNextNonBlocked(BarrierTracker.java:94) at org.apache.flink.streaming.runtime.io.StreamInputProcessor.processInput(StreamInputProcessor.java:209) at org.apache.flink.streaming.runtime.tasks.OneInputStreamTask.run(OneInputStreamTask.java:103) at org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:306) at org.apache.flink.runtime.taskmanager.Task.run(Task.java:703) at java.lang.Thread.run(Thread.java:748) 这个线程对应了我们sql里的union算子 --- "groupBy: (pro, throwable, level, ip), window: (TumblingGroupWindow('w$, 'SPT, 3000.millis)), select: (pro, throwable, level, ip, COUNT(*) AS count, lastStrInGroupSkipNull($f5) AS id, firstLong(l) AS firstl, lastLong(l) AS lastl, start('w$) AS w$start, end('w$) AS w$end, rowtime('w$) AS w$rowtime, proctime('w$) AS w$proctime) -> where: (=(1, uniqueWithin100MS(pro, throwable, _UTF-16LE'ERROR', ip, w$end))), select: (pro, throwable, level, ip, count, id, w$end AS time, firstl, lastl) -> to: Row -> Sink: Kafka010JsonTableSink(pro, throwable, level, ip, count, id, time, firstl, lastl) (1/1)" #54 prio=5 os_prio=0 tid=0x00007f363fde3800 nid=0x664 in Object.wait() [0x00007f3641127000] java.lang.Thread.State: WAITING (on object monitor) at java.lang.Object.wait(Native Method) at java.lang.Object.wait(Object.java:502) at org.apache.flink.runtime.io.network.partition.consumer.SingleInputGate.getNextBufferOrEvent(SingleInputGate.java:533) - locked <0x00000000e6ee2d48> (a java.util.ArrayDeque) at org.apache.flink.runtime.io.network.partition.consumer.SingleInputGate.getNextBufferOrEvent(SingleInputGate.java:502) at org.apache.flink.streaming.runtime.io.BarrierTracker.getNextNonBlocked(BarrierTracker.java:94) at org.apache.flink.streaming.runtime.io.StreamInputProcessor.processInput(StreamInputProcessor.java:209) at org.apache.flink.streaming.runtime.tasks.OneInputStreamTask.run(OneInputStreamTask.java:103) at org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:306) at org.apache.flink.runtime.taskmanager.Task.run(Task.java:703) at java.lang.Thread.run(Thread.java:748) 这个对应了group by算子 ---生产者 "kafka-producer-network-thread | producer-1" #55 daemon prio=5 os_prio=0 tid=0x00007f364d0f0800 nid=0x667 runnable [0x00007f3640a26000] java.lang.Thread.State: RUNNABLE at sun.nio.ch.EPollArrayWrapper.epollWait(Native Method) at sun.nio.ch.EPollArrayWrapper.poll(EPollArrayWrapper.java:269) at sun.nio.ch.EPollSelectorImpl.doSelect(EPollSelectorImpl.java:93) at sun.nio.ch.SelectorImpl.lockAndDoSelect(SelectorImpl.java:86) - locked <0x00000000e6ef3358> (a sun.nio.ch.Util$3) - locked <0x00000000e6ef3340> (a java.util.Collections$UnmodifiableSet) - locked <0x00000000e6eedbd8> (a sun.nio.ch.EPollSelectorImpl) at sun.nio.ch.SelectorImpl.select(SelectorImpl.java:97) at org.apache.kafka.common.network.Selector.select(Selector.java:489) at org.apache.kafka.common.network.Selector.poll(Selector.java:298) at org.apache.kafka.clients.NetworkClient.poll(NetworkClient.java:349) at org.apache.kafka.clients.producer.internals.Sender.run(Sender.java:225) at org.apache.kafka.clients.producer.internals.Sender.run(Sender.java:126) at java.lang.Thread.run(Thread.java:748) 对应着生产者，直连kafka --- "Time Trigger for Source: MyKafka010JsonTableSource -> from: (l, KAFKA_TOPIC, KAFKA_PARTITION, KAFKA_OFFSET, pro, throwable, level, ip, SPT) -> Timestamps/Watermarks -> where: (AND(=(level, _UTF-16LE'ERROR'), IS NOT NULL(pro), IS NOT NULL(throwable), IS NOT NULL(ip))), select: (pro, throwable, CAST(_UTF-16LE'ERROR') AS level, ip, SPT, CONCAT_WS(_UTF-16LE'_', KAFKA_TOPIC, CAST(KAFKA_PARTITION), CAST(KAFKA_OFFSET)) AS $f5, l) (1/1)" #57 daemon prio=5 os_prio=0 tid=0x00007f364d264800 nid=0x669 waiting on condition [0x00007f3640624000] java.lang.Thread.State: TIMED_WAITING (parking) at sun.misc.Unsafe.park(Native Method) - parking to wait for <0x00000000e6ef84c0> (a java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject) at java.util.concurrent.locks.LockSupport.parkNanos(LockSupport.java:215) at java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.awaitNanos(AbstractQueuedSynchronizer.java:2078) at java.util.concurrent.ScheduledThreadPoolExecutor$DelayedWorkQueue.take(ScheduledThreadPoolExecutor.java:1093) at java.util.concurrent.ScheduledThreadPoolExecutor$DelayedWorkQueue.take(ScheduledThreadPoolExecutor.java:809) at java.util.concurrent.ThreadPoolExecutor.getTask(ThreadPoolExecutor.java:1067) at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1127) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617) at java.lang.Thread.run(Thread.java:748) "Time Trigger for Source: MyKafka010JsonTableSource -> from: (l, KAFKA_TOPIC, KAFKA_PARTITION, KAFKA_OFFSET, pro, throwable, level, ip, SPT) -> Timestamps/Watermarks -> where: (AND(=(level, _UTF-16LE'ERROR'), IS NOT NULL(pro), IS NOT NULL(throwable), IS NOT NULL(ip))), select: (pro, throwable, CAST(_UTF-16LE'ERROR') AS level, ip, SPT, CONCAT_WS(_UTF-16LE'_', KAFKA_TOPIC, CAST(KAFKA_PARTITION), CAST(KAFKA_OFFSET)) AS $f5, l) (1/1)" #56 daemon prio=5 os_prio=0 tid=0x00007f363e937800 nid=0x668 waiting on condition [0x00007f3640725000] java.lang.Thread.State: TIMED_WAITING (parking) at sun.misc.Unsafe.park(Native Method) - parking to wait for <0x00000000e6ee2bc8> (a java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject) at java.util.concurrent.locks.LockSupport.parkNanos(LockSupport.java:215) at java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.awaitNanos(AbstractQueuedSynchronizer.java:2078) at java.util.concurrent.ScheduledThreadPoolExecutor$DelayedWorkQueue.take(ScheduledThreadPoolExecutor.java:1093) at java.util.concurrent.ScheduledThreadPoolExecutor$DelayedWorkQueue.take(ScheduledThreadPoolExecutor.java:809) at java.util.concurrent.ThreadPoolExecutor.getTask(ThreadPoolExecutor.java:1067) at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1127) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617) at java.lang.Thread.run(Thread.java:748) 每个流对应着一个水印定时发送线程，因为我这边的输入是2个流 所以有2个水印发送线程 --- "Kafka Partition Discovery for Source: MyKafka010JsonTableSource -> from: (l, KAFKA_TOPIC, KAFKA_PARTITION, KAFKA_OFFSET, pro, throwable, level, ip, SPT) -> Timestamps/Watermarks -> where: (AND(=(level, _UTF-16LE'ERROR'), IS NOT NULL(pro), IS NOT NULL(throwable), IS NOT NULL(ip))), select: (pro, throwable, CAST(_UTF-16LE'ERROR') AS level, ip, SPT, CONCAT_WS(_UTF-16LE'_', KAFKA_TOPIC, CAST(KAFKA_PARTITION), CAST(KAFKA_OFFSET)) AS $f5, l) (1/1)" #61 prio=5 os_prio=0 tid=0x00007f364d25f000 nid=0x66c waiting on condition [0x00007f3640121000] java.lang.Thread.State: TIMED_WAITING (sleeping) at java.lang.Thread.sleep(Native Method) at org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumerBase$2.run(FlinkKafkaConsumerBase.java:701) at java.lang.Thread.run(Thread.java:748) "Kafka Partition Discovery for Source: MyKafka010JsonTableSource -> from: (l, KAFKA_TOPIC, KAFKA_PARTITION, KAFKA_OFFSET, pro, throwable, level, ip, SPT) -> Timestamps/Watermarks -> where: (AND(=(level, _UTF-16LE'ERROR'), IS NOT NULL(pro), IS NOT NULL(throwable), IS NOT NULL(ip))), select: (pro, throwable, CAST(_UTF-16LE'ERROR') AS level, ip, SPT, CONCAT_WS(_UTF-16LE'_', KAFKA_TOPIC, CAST(KAFKA_PARTITION), CAST(KAFKA_OFFSET)) AS $f5, l) (1/1)" #59 prio=5 os_prio=0 tid=0x00007f363f4bc800 nid=0x66a waiting on condition [0x00007f3640323000] java.lang.Thread.State: TIMED_WAITING (sleeping) at java.lang.Thread.sleep(Native Method) at org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumerBase$2.run(FlinkKafkaConsumerBase.java:701) at java.lang.Thread.run(Thread.java:748) 2个自动分区发现线程 --- "Kafka 0.10 Fetcher for Source: MyKafka010JsonTableSource -> from: (l, KAFKA_TOPIC, KAFKA_PARTITION, KAFKA_OFFSET, pro, throwable, level, ip, SPT) -> Timestamps/Watermarks -> where: (AND(=(level, _UTF-16LE'ERROR'), IS NOT NULL(pro), IS NOT NULL(throwable), IS NOT NULL(ip))), select: (pro, throwable, CAST(_UTF-16LE'ERROR') AS level, ip, SPT, CONCAT_WS(_UTF-16LE'_', KAFKA_TOPIC, CAST(KAFKA_PARTITION), CAST(KAFKA_OFFSET)) AS $f5, l) (1/1)" #60 daemon prio=5 os_prio=0 tid=0x00007f364d269800 nid=0x66d runnable [0x00007f363bffe000] java.lang.Thread.State: RUNNABLE at sun.nio.ch.EPollArrayWrapper.epollWait(Native Method) at sun.nio.ch.EPollArrayWrapper.poll(EPollArrayWrapper.java:269) at sun.nio.ch.EPollSelectorImpl.doSelect(EPollSelectorImpl.java:93) at sun.nio.ch.SelectorImpl.lockAndDoSelect(SelectorImpl.java:86) - locked <0x00000000e73f0888> (a sun.nio.ch.Util$3) - locked <0x00000000e73f0870> (a java.util.Collections$UnmodifiableSet) - locked <0x00000000e7279b20> (a sun.nio.ch.EPollSelectorImpl) at sun.nio.ch.SelectorImpl.select(SelectorImpl.java:97) at org.apache.kafka.common.network.Selector.select(Selector.java:489) at org.apache.kafka.common.network.Selector.poll(Selector.java:298) at org.apache.kafka.clients.NetworkClient.poll(NetworkClient.java:349) at org.apache.kafka.clients.consumer.internals.ConsumerNetworkClient.poll(ConsumerNetworkClient.java:226) - locked <0x00000000e7497ec0> (a org.apache.kafka.clients.consumer.internals.ConsumerNetworkClient) at org.apache.kafka.clients.consumer.KafkaConsumer.pollOnce(KafkaConsumer.java:1047) at org.apache.kafka.clients.consumer.KafkaConsumer.poll(KafkaConsumer.java:995) at org.apache.flink.streaming.connectors.kafka.internal.KafkaConsumerThread.run(KafkaConsumerThread.java:257) "Kafka 0.10 Fetcher for Source: MyKafka010JsonTableSource -> from: (l, KAFKA_TOPIC, KAFKA_PARTITION, KAFKA_OFFSET, pro, throwable, level, ip, SPT) -> Timestamps/Watermarks -> where: (AND(=(level, _UTF-16LE'ERROR'), IS NOT NULL(pro), IS NOT NULL(throwable), IS NOT NULL(ip))), select: (pro, throwable, CAST(_UTF-16LE'ERROR') AS level, ip, SPT, CONCAT_WS(_UTF-16LE'_', KAFKA_TOPIC, CAST(KAFKA_PARTITION), CAST(KAFKA_OFFSET)) AS $f5, l) (1/1)" #58 daemon prio=5 os_prio=0 tid=0x00007f363f4be800 nid=0x66b runnable [0x00007f3640222000] java.lang.Thread.State: RUNNABLE at sun.nio.ch.EPollArrayWrapper.epollWait(Native Method) at sun.nio.ch.EPollArrayWrapper.poll(EPollArrayWrapper.java:269) at sun.nio.ch.EPollSelectorImpl.doSelect(EPollSelectorImpl.java:93) at sun.nio.ch.SelectorImpl.lockAndDoSelect(SelectorImpl.java:86) - locked <0x00000000e6ef0758> (a sun.nio.ch.Util$3) - locked <0x00000000e6ef0740> (a java.util.Collections$UnmodifiableSet) - locked <0x00000000e6ee0248> (a sun.nio.ch.EPollSelectorImpl) at sun.nio.ch.SelectorImpl.select(SelectorImpl.java:97) at org.apache.kafka.common.network.Selector.select(Selector.java:489) at org.apache.kafka.common.network.Selector.poll(Selector.java:298) at org.apache.kafka.clients.NetworkClient.poll(NetworkClient.java:349) at org.apache.kafka.clients.consumer.internals.ConsumerNetworkClient.poll(ConsumerNetworkClient.java:226) - locked <0x00000000e6f03398> (a org.apache.kafka.clients.consumer.internals.ConsumerNetworkClient) at org.apache.kafka.clients.consumer.KafkaConsumer.pollOnce(KafkaConsumer.java:1047) at org.apache.kafka.clients.consumer.KafkaConsumer.poll(KafkaConsumer.java:995) at org.apache.flink.streaming.connectors.kafka.internal.KafkaConsumerThread.run(KafkaConsumerThread.java:257) 对应着2个直连kafka的生产者线程 线程debug完了，下面我们来看每个线程做什么事情！这里先简单交代一下消息记录和watermark的背景 对于每个流，有1个消费者线程来读取kafka的消息 然后通过本地内存交换，喂给另外一个线程，就是文中Handover字样的线程，这个线程会把消息往下游发送，同时，有1个水印线程定时探测是否有更大时间戳出现，出现的话，把这个时间戳放在一个水印事件里下广播给下游. ---下面先来debug下Handover线程，看看是如何消息喂给unionInputGate线程的 断点在 stop at org.apache.flink.streaming.connectors.kafka.internal.Kafka09Fetcher:154 跑起来！ 然后，发送一条消息到kafka,断点顺利命中 接下来就是具体看消息的流转过程！ 消息处理过程中，会记录下当前事件的时间戳，位置在 作用是如果时间戳比当前值更大，则更新这个时间戳，后面会有水印线程定时读取这个值决定是否需要发送水印信息 好，继续观察消息的流动，执行到了下面这个地方 [1] org.apache.flink.runtime.io.network.api.writer.RecordWriter.emit (RecordWriter.java:104) [2] org.apache.flink.streaming.runtime.io.StreamRecordWriter.emit (StreamRecordWriter.java:81) [3] org.apache.flink.streaming.runtime.io.RecordWriterOutput.pushToRecordWriter (RecordWriterOutput.java:107) [4] org.apache.flink.streaming.runtime.io.RecordWriterOutput.collect (RecordWriterOutput.java:89) [5] org.apache.flink.streaming.runtime.io.RecordWriterOutput.collect (RecordWriterOutput.java:45) [6] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:679) [7] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:657) [8] org.apache.flink.streaming.api.operators.TimestampedCollector.collect (TimestampedCollector.java:51) [9] org.apache.flink.table.runtime.CRowWrappingCollector.collect (CRowWrappingCollector.scala:37) [10] org.apache.flink.table.runtime.CRowWrappingCollector.collect (CRowWrappingCollector.scala:28) [11] DataStreamCalcRule$69.processElement (null) [12] org.apache.flink.table.runtime.CRowProcessRunner.processElement (CRowProcessRunner.scala:66) [13] org.apache.flink.table.runtime.CRowProcessRunner.processElement (CRowProcessRunner.scala:35) [14] org.apache.flink.streaming.api.operators.ProcessOperator.processElement (ProcessOperator.java:66) [15] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.pushToOperator (OperatorChain.java:560) [16] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect (OperatorChain.java:535) [17] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect (OperatorChain.java:515) [18] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:679) [19] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:657) [20] org.apache.flink.streaming.runtime.operators.TimestampsAndPeriodicWatermarksOperator.processElement (TimestampsAndPeriodicWatermarksOperator.java:67) [21] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.pushToOperator (OperatorChain.java:560) [22] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect (OperatorChain.java:535) [23] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect (OperatorChain.java:515) [24] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:679) [25] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:657) [26] org.apache.flink.streaming.api.operators.TimestampedCollector.collect (TimestampedCollector.java:51) [27] org.apache.flink.table.runtime.CRowWrappingCollector.collect (CRowWrappingCollector.scala:37) [28] org.apache.flink.table.runtime.CRowWrappingCollector.collect (CRowWrappingCollector.scala:28) [29] DataStreamSourceConversion$23.processElement (null) [30] org.apache.flink.table.runtime.CRowOutputProcessRunner.processElement (CRowOutputProcessRunner.scala:67) [31] org.apache.flink.streaming.api.operators.ProcessOperator.processElement (ProcessOperator.java:66) [32] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.pushToOperator (OperatorChain.java:560) [33] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect (OperatorChain.java:535) [34] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect (OperatorChain.java:515) [35] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:679) [36] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:657) [37] org.apache.flink.streaming.api.operators.StreamSourceContexts$ManualWatermarkContext.processAndCollectWithTimestamp (StreamSourceContexts.java:310) [38] org.apache.flink.streaming.api.operators.StreamSourceContexts$WatermarkContext.collectWithTimestamp (StreamSourceContexts.java:409) [39] org.apache.flink.streaming.connectors.kafka.internals.AbstractFetcher.emitRecordWithTimestamp (AbstractFetcher.java:398) [40] org.apache.flink.streaming.connectors.kafka.internal.Kafka010Fetcher.emitRecord (Kafka010Fetcher.java:89) [41] org.apache.flink.streaming.connectors.kafka.internal.Kafka09Fetcher.runFetchLoop (Kafka09Fetcher.java:154) [42] org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumerBase.run (FlinkKafkaConsumerBase.java:721) [43] org.apache.flink.streaming.api.operators.StreamSource.run (StreamSource.java:87) [44] org.apache.flink.streaming.api.operators.StreamSource.run (StreamSource.java:56) [45] org.apache.flink.streaming.runtime.tasks.SourceStreamTask.run (SourceStreamTask.java:99) [46] org.apache.flink.streaming.runtime.tasks.StreamTask.invoke (StreamTask.java:306) [47] org.apache.flink.runtime.taskmanager.Task.run (Task.java:703) [48] java.lang.Thread.run (Thread.java:748) 看一下这里的即将执行的代码 public void emit(T record) throws IOException, InterruptedException { for (int targetChannel : channelSelector.selectChannels(record, numChannels)) { sendToTarget(record, targetChannel); } } 这里的print numChannels numChannels = 1 --->因为我们有一个union操作，union自然是所有源归一！这就对了！ ---最后放入消息并提醒消费线程，完整的调用栈如下： [1] org.apache.flink.runtime.io.network.partition.consumer.SingleInputGate.queueChannel (SingleInputGate.java:623) [2] org.apache.flink.runtime.io.network.partition.consumer.SingleInputGate.notifyChannelNonEmpty (SingleInputGate.java:612) [3] org.apache.flink.runtime.io.network.partition.consumer.InputChannel.notifyChannelNonEmpty (InputChannel.java:121) [4] org.apache.flink.runtime.io.network.partition.consumer.LocalInputChannel.notifyDataAvailable (LocalInputChannel.java:202) [5] org.apache.flink.runtime.io.network.partition.PipelinedSubpartitionView.notifyDataAvailable (PipelinedSubpartitionView.java:56) [6] org.apache.flink.runtime.io.network.partition.PipelinedSubpartition.notifyDataAvailable (PipelinedSubpartition.java:290) [7] org.apache.flink.runtime.io.network.partition.PipelinedSubpartition.flush (PipelinedSubpartition.java:76) [8] org.apache.flink.runtime.io.network.partition.ResultPartition.flush (ResultPartition.java:269) [9] org.apache.flink.runtime.io.network.api.writer.RecordWriter.sendToTarget (RecordWriter.java:149) [10] org.apache.flink.runtime.io.network.api.writer.RecordWriter.emit (RecordWriter.java:105) [11] org.apache.flink.streaming.runtime.io.StreamRecordWriter.emit (StreamRecordWriter.java:81) [12] org.apache.flink.streaming.runtime.io.RecordWriterOutput.pushToRecordWriter (RecordWriterOutput.java:107) [13] org.apache.flink.streaming.runtime.io.RecordWriterOutput.collect (RecordWriterOutput.java:89) [14] org.apache.flink.streaming.runtime.io.RecordWriterOutput.collect (RecordWriterOutput.java:45) [15] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:679) [16] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:657) [17] org.apache.flink.streaming.api.operators.TimestampedCollector.collect (TimestampedCollector.java:51) [18] org.apache.flink.table.runtime.CRowWrappingCollector.collect (CRowWrappingCollector.scala:37) [19] org.apache.flink.table.runtime.CRowWrappingCollector.collect (CRowWrappingCollector.scala:28) [20] DataStreamCalcRule$69.processElement (null) [21] org.apache.flink.table.runtime.CRowProcessRunner.processElement (CRowProcessRunner.scala:66) [22] org.apache.flink.table.runtime.CRowProcessRunner.processElement (CRowProcessRunner.scala:35) [23] org.apache.flink.streaming.api.operators.ProcessOperator.processElement (ProcessOperator.java:66) [24] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.pushToOperator (OperatorChain.java:560) [25] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect (OperatorChain.java:535) [26] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect (OperatorChain.java:515) [27] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:679) [28] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:657) [29] org.apache.flink.streaming.runtime.operators.TimestampsAndPeriodicWatermarksOperator.processElement (TimestampsAndPeriodicWatermarksOperator.java:67) [30] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.pushToOperator (OperatorChain.java:560) [31] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect (OperatorChain.java:535) [32] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect (OperatorChain.java:515) [33] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:679) [34] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:657) [35] org.apache.flink.streaming.api.operators.TimestampedCollector.collect (TimestampedCollector.java:51) [36] org.apache.flink.table.runtime.CRowWrappingCollector.collect (CRowWrappingCollector.scala:37) [37] org.apache.flink.table.runtime.CRowWrappingCollector.collect (CRowWrappingCollector.scala:28) [38] DataStreamSourceConversion$23.processElement (null) [39] org.apache.flink.table.runtime.CRowOutputProcessRunner.processElement (CRowOutputProcessRunner.scala:67) [40] org.apache.flink.streaming.api.operators.ProcessOperator.processElement (ProcessOperator.java:66) [41] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.pushToOperator (OperatorChain.java:560) [42] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect (OperatorChain.java:535) [43] org.apache.flink.streaming.runtime.tasks.OperatorChain$CopyingChainingOutput.collect (OperatorChain.java:515) [44] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:679) [45] org.apache.flink.streaming.api.operators.AbstractStreamOperator$CountingOutput.collect (AbstractStreamOperator.java:657) [46] org.apache.flink.streaming.api.operators.StreamSourceContexts$ManualWatermarkContext.processAndCollectWithTimestamp (StreamSourceContexts.java:310) [47] org.apache.flink.streaming.api.operators.StreamSourceContexts$WatermarkContext.collectWithTimestamp (StreamSourceContexts.java:409) [48] org.apache.flink.streaming.connectors.kafka.internals.AbstractFetcher.emitRecordWithTimestamp (AbstractFetcher.java:398) [49] org.apache.flink.streaming.connectors.kafka.internal.Kafka010Fetcher.emitRecord (Kafka010Fetcher.java:89) [50] org.apache.flink.streaming.connectors.kafka.internal.Kafka09Fetcher.runFetchLoop (Kafka09Fetcher.java:154) [51] org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumerBase.run (FlinkKafkaConsumerBase.java:721) [52] org.apache.flink.streaming.api.operators.StreamSource.run (StreamSource.java:87) [53] org.apache.flink.streaming.api.operators.StreamSource.run (StreamSource.java:56) [54] org.apache.flink.streaming.runtime.tasks.SourceStreamTask.run (SourceStreamTask.java:99) [55] org.apache.flink.streaming.runtime.tasks.StreamTask.invoke (StreamTask.java:306) [56] org.apache.flink.runtime.taskmanager.Task.run (Task.java:703) [57] java.lang.Thread.run (Thread.java:748) ---水印的处理应该也是类似的，所以接下来，我们来看Union所在的线程 我们再来复习下上面里提到的这个线程的调用栈 "time attribute: (SPT) (1/1)" #53 prio=5 os_prio=0 tid=0x00007f363d8e4000 nid=0x662 in Object.wait() [0x00007f36418c1000] java.lang.Thread.State: WAITING (on object monitor) at java.lang.Object.wait(Native Method) at java.lang.Object.wait(Object.java:502) at org.apache.flink.runtime.io.network.partition.consumer.UnionInputGate.waitAndGetNextInputGate(UnionInputGate.java:205) - locked <0x00000000e6ee8210> (a java.util.ArrayDeque) at org.apache.flink.runtime.io.network.partition.consumer.UnionInputGate.getNextBufferOrEvent(UnionInputGate.java:163) at org.apache.flink.streaming.runtime.io.BarrierTracker.getNextNonBlocked(BarrierTracker.java:94) at org.apache.flink.streaming.runtime.io.StreamInputProcessor.processInput(StreamInputProcessor.java:209) at org.apache.flink.streaming.runtime.tasks.OneInputStreamTask.run(OneInputStreamTask.java:103) at org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:306) at org.apache.flink.runtime.taskmanager.Task.run(Task.java:703) at java.lang.Thread.run(Thread.java:748) 这个线程对应了我们sql里的union算子 上面这个图，是等待有消息过来就提取消息，任何一个源有消息都会触发消息提取，否则wait ---注意：这里的消息有4种类型，一般我们只需要关注record+watermark即可 具体地点是： ---这里讲一下，关于LatencyMarker，默认2秒钟发送一次，截图如下 其它的不管是record还是watermark都会往下发送！ 下面我们来在union里同时针对record和watermark打断点，猜一猜哪个断点先被触发？ 断点位于【针对flink-1.5版本】 Breakpoints set: breakpoint org.apache.flink.streaming.runtime.io.StreamInputProcessor:184 breakpoint org.apache.flink.streaming.runtime.io.StreamInputProcessor:198 触发的顺序如下： ---跟想的是一样的！ 下面就去研究下groupby线程 "groupBy: (pro, throwable, level, ip), window: (TumblingGroupWindow('w$, 'SPT, 3000.millis)), select: (pro, throwable, level, ip, COUNT(*) AS count, lastStrInGroupSkipNull($f5) AS id, firstLong(l) AS firstl, lastLong(l) AS lastl, start('w$) AS w$start, end('w$) AS w$end, rowtime('w$) AS w$rowtime, proctime('w$) AS w$proctime) -> where: (=(1, uniqueWithin100MS(pro, throwable, _UTF-16LE'ERROR', ip, w$end))), select: (pro, throwable, level, ip, count, id, w$end AS time, firstl, lastl) -> to: Row -> Sink: Kafka010JsonTableSink(pro, throwable, level, ip, count, id, time, firstl, lastl) (1/1)" #54 prio=5 os_prio=0 tid=0x00007f363fde3800 nid=0x664 in Object.wait() [0x00007f3641127000] java.lang.Thread.State: WAITING (on object monitor) at java.lang.Object.wait(Native Method) at java.lang.Object.wait(Object.java:502) at org.apache.flink.runtime.io.network.partition.consumer.SingleInputGate.getNextBufferOrEvent(SingleInputGate.java:533) - locked <0x00000000e6ee2d48> (a java.util.ArrayDeque) at org.apache.flink.runtime.io.network.partition.consumer.SingleInputGate.getNextBufferOrEvent(SingleInputGate.java:502) at org.apache.flink.streaming.runtime.io.BarrierTracker.getNextNonBlocked(BarrierTracker.java:94) at org.apache.flink.streaming.runtime.io.StreamInputProcessor.processInput(StreamInputProcessor.java:209) at org.apache.flink.streaming.runtime.tasks.OneInputStreamTask.run(OneInputStreamTask.java:103) at org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:306) at org.apache.flink.runtime.taskmanager.Task.run(Task.java:703) at java.lang.Thread.run(Thread.java:748) 这个对应了group by算子 针对group by来说，最重要的环节，这个其实跟union线程一样的，也是在 org.apache.flink.streaming.runtime.io.StreamInputProcessor.processInput 这里面来做事件的分发，所以断点都是一样的 --- 这里主要强调，在groupby处理watermark时的位置如下：【尤其是针对多个source来说，很容易出问题】 这个时候，我意识到在groupby线程中来观察watermark还早了点，因为在union线程中针对watermark的处理还有一些秘密 所以我们回到union线程来挖这些秘密，把groupby线程用suspend命令挂起来，专门debug union线程即可！ ---打个断点【针对flink-1.5】 stop at org.apache.flink.streaming.runtime.io.StreamInputProcessor:184 研究了一把，大致明白原理了，这么说吧，线程模型如下 流1------- | | | | | |---------->union线程的watermark--------->groupby线程的watermark | | | | 流2------- 其中，流1和流2---每次都发送自己看到的最大时间戳发送个下游(看到小的就什么都不做) union这里会动态更新流1和流2的各自所看到的最大时间戳，同时取Min（流1的最大时间戳，流2的最大时间戳），跟上一次的值比较 如果>上一次的Min值，则发送给group by. ---我觉得读者看到这里，肯定已经懵逼了，我来解释下思想 强调一下：消息在中间过程中不拦截，直达最后的windowoperator那里做windowLate判断决定是否丢弃！ =========================================================================================== 对于流1来说，它每次发送自己已知的最大时间戳给下游，就是说“你好，下游，对我来说小于这个时间戳的就算是延迟消息，你看着办” 对于流2来说，它每次发送自己已知的最大时间戳给下游，就是说“你好，下游，对我来说小于这个时间戳的就算是延迟消息，你看着办” ---对于union来说，这里复杂些 它取值min( 流1的max时间戳，流2的max时间戳）跟上一次的min( 流1的max时间戳，流2的max时间戳）比较， 如果发现递增了，就把当前较大的这个min值发送给下游，说“你好，下游，全局来说，对我来说小于这个时间戳的就算是延迟消息，我只能帮到这里了，已经尽力拖住时间戳了,你看着办” ---对于groupby来说，它收到时间戳，每次保留最大值，然后参考最大值来快速决定每个消息是不是延迟消息(最大值-可容忍的延迟消息)。 所以，在多源情况下，判断全局一个消息是不是延迟消息，实际上由min( 流1的max时间戳，流2的max时间戳）这个值来参与决定 --- 我们再跳出来想一想这个事情，我估计读者最懵逼的地方就是union为啥取每个流的最小值，而不是最大值 我们就这么理解吧，如果取最大值，那消费慢的流的数据大部分都成为了late数据被丢弃，union就会被打 所以union为了防止被打，它不想惹众怒，就取了min(每个流），这样所有人都无话可说了 union旁白：我都取了你们每个流的各自的时间戳最大值的全局最小值，还要我怎么样， 最慢的那个流也不会说啥了，因为取的就是它这个流上报的自身最大值。 上面都是从技术角度来阐述这个事情，那么我们再拔高一下，从更高的层次来看这个事情 其实就是让更多的数据没有成为late数据，纳入正常运算范围内，由min( 流1的max时间戳，流2的max时间戳）的递增来推动全局windowoperator的计算输出结果. 相应的，消费最慢的流会拖累最终业务数据的延迟生成. ---读者可以再细细琢磨里面的门道，下面我们来做逻辑测试！验证我们是否真正理解了这个游戏规则! 背景：容忍延迟3000毫秒 下面每行的格式就是：流名称 + 时间戳 ，每次只输出1条 1)流1 + 1545703896000 2)流1 + 1545703896000 3)流2 + 1545703896000 4)流2 + 1545703898999 5)流2 + 1545703899000 6)流1 + 1545703899000 7)流1 + 1545703900000 8)流2 + 1545703902000-1 --->这个不会触发windowOperator的输出，因为流1的最小值还不够 9)流1 + 1545703902000-1 --->这个才会触发windowOperator的输出 正确输出了，记住，一定要2个流 【齐头并进，理实交融】 但是，其实，仅仅研究到这一步，并没有完全结束，欲知后事如何请听下回分解 :)
