DDR爱好者之家 Design By 杰米
在采用log4j的kafka-appender收集spark任务运行日志时，发现提交到yarn上的任务始终ACCEPTED状态，无法进入RUNNING状态，并且会重试两次后超时。期初认为是yarn资源不足导致，但在确认yarn资源充裕的时候问题依旧，而且基本上能稳定复现。
起初是这么配置spark日志输出到kafka的：
log4j.rootCategory=INFO, console, kafka
log4j.appender.console=org.apache.log4j.ConsoleAppender
log4j.appender.console.target=System.err
log4j.appender.console.layout=org.apache.log4j.PatternLayout
log4j.appender.console.layout.ConversionPattern=%d{yyyy/MM/dd HH:mm:ss.SSS} %p %c{1}: [${log4j.pipelineId}] %m%n
# Kafka appender
log4j.appender.kafka=org.apache.kafka.log4jappender.KafkaLog4jAppender
# Set Kafka topic and brokerList
log4j.appender.kafka.topic=yarn_spark_log
log4j.appender.kafka.brokerList=localhost:9092
log4j.appender.kafka.compressionType=none
log4j.appender.kafka.syncSend=false
log4j.appender.kafka.maxBlockMs=10
log4j.appender.kafka.layout=org.apache.log4j.PatternLayout
log4j.appender.kafka.layout.ConversionPattern=%d{yyyy/MM/dd HH:mm:ss.SSS} %p %c{1}: [${log4j.pipelineId}] %m
这里用org.apache.kafka.log4jappender.KafkaLog4jAppender默认将所有日志都输出到kafka，这个appender已经被kafka官方维护，稳定性应该是可以保障的。
问题定位
发现问题后，尝试将输出到kafka的规则去掉，问题解除！于是把问题定位到跟日志输出到kafka有关。通过其他测试，证实目标kafka其实是正常的，这就非常奇怪了。
查看yarn的ResourceManager日志，发现有如下超时
2020-05-07 21:49:48,230 INFO org.apache.hadoop.yarn.util.AbstractLivelinessMonitor: Expired:appattempt_1578970174552_3204_000002 Timed out after 600 secs
2020-05-07 21:49:48,230 INFO org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptImpl: Updating application attempt appattempt_1578970174552_3204_000002 with final
state: FAILED, and exit status: -1000
2020-05-07 21:49:48,231 INFO org.apache.hadoop.yarn.server.resourcemanager.rmapp.attempt.RMAppAttemptImpl: appattempt_1578970174552_3204_000002 State change from LAUNCHED to FINAL_SAV
ING on event = EXPIRE
表明，yarn本身是接收任务的，但是发现任务迟迟没有启动。在spark的场景下其实是指只有driver启动了，但是没有启动executor。
而查看driver日志，发现日志输出到一个地方就卡住了，不往下继续了。通过对比成功运行和卡住的情况发现，日志卡在这条上：
2020/05/07 19:37:10.324 INFO SecurityManager: Changing view acls to: yarn,root
2020/05/07 19:37:10.344 INFO Metadata: Cluster ID: 6iG6WHA2SoK7FfgGgWHt_A
卡住的情况下，只会打出SecurityManager这行，而无法打出Metadata这行。
猜想Metadata这行是kafka-client本身打出来的，因为整个上下文只有yarn, spark, kafka-client可能会打出这个日志。
在kafka-client 2.2.0版本中找到这个日志是输出位置：
public synchronized void update(MetadataResponse metadataResponse, long now) {
...
String newClusterId = cache.cluster().clusterResource().clusterId();
if (!Objects.equals(previousClusterId, newClusterId)) {
log.info("Cluster ID: {}", newClusterId);
}
...
}
看到synchronized，高度怀疑死锁。于是考虑用jstack分析：
在yarn上运行spark任务的时候，driver进程叫ApplicationMaster，executor进程叫CoarseGrainedExecutorBackend。这里首先尝试再复现过程中找到drvier最终在哪个节点上运行，然后快速使用jstack -F <pid>打印堆栈
jstack果然不负众望，报告了死锁！这里我把结果贴的全一点
[root@node1 ~]# jstack 20136 20136: Unable to open socket file: target process not responding or HotSpot VM not loaded The -F option can be used when the target process is not responding [root@node1 ~]# jstack -F 20136 Attaching to process ID 20136, please wait... Debugger attached successfully. Server compiler detected. JVM version is 25.231-b11 Deadlock Detection: Found one Java-level deadlock: ============================= "kafka-producer-network-thread | producer-1": waiting to lock Monitor@0x00000000025fcc48 (Object@0x00000000ed680b60, a org/apache/kafka/log4jappender/KafkaLog4jAppender), which is held by "main" "main": waiting to lock Monitor@0x00007fec9dbde038 (Object@0x00000000ee44de38, a org/apache/kafka/clients/Metadata), which is held by "kafka-producer-network-thread | producer-1" Found a total of 1 deadlock. Thread 20157: (state = BLOCKED) - org.apache.log4j.AppenderSkeleton.doAppend(org.apache.log4j.spi.LoggingEvent) @bci=0, line=231 (Interpreted frame) - org.apache.log4j.helpers.AppenderAttachableImpl.appendLoopOnAppenders(org.apache.log4j.spi.LoggingEvent) @bci=41, line=66 (Interpreted frame) - org.apache.log4j.Category.callAppenders(org.apache.log4j.spi.LoggingEvent) @bci=26, line=206 (Interpreted frame) - org.apache.log4j.Category.forcedLog(java.lang.String, org.apache.log4j.Priority, java.lang.Object, java.lang.Throwable) @bci=14, line=391 (Interpreted frame) - org.apache.log4j.Category.log(java.lang.String, org.apache.log4j.Priority, java.lang.Object, java.lang.Throwable) @bci=34, line=856 (Interpreted frame) - org.slf4j.impl.Log4jLoggerAdapter.info(java.lang.String, java.lang.Object) @bci=34, line=324 (Interpreted frame) - org.apache.kafka.clients.Metadata.update(org.apache.kafka.common.requests.MetadataResponse, long) @bci=317, line=365 (Interpreted frame) - org.apache.kafka.clients.NetworkClient$DefaultMetadataUpdater.handleCompletedMetadataResponse(org.apache.kafka.common.requests.RequestHeader, long, org.apache.kafka.common.requests.MetadataResponse) @bci=184, line=1031 (Interpreted frame) - org.apache.kafka.clients.NetworkClient.handleCompletedReceives(java.util.List, long) @bci=215, line=822 (Interpreted frame) - org.apache.kafka.clients.NetworkClient.poll(long, long) @bci=132, line=544 (Interpreted frame) - org.apache.kafka.clients.producer.internals.Sender.run(long) @bci=227, line=311 (Interpreted frame) - org.apache.kafka.clients.producer.internals.Sender.run() @bci=28, line=235 (Interpreted frame) - java.lang.Thread.run() @bci=11, line=748 (Interpreted frame) Thread 20150: (state = BLOCKED) Thread 20149: (state = BLOCKED) - java.lang.Object.wait(long) @bci=0 (Interpreted frame) - java.lang.ref.ReferenceQueue.remove(long) @bci=59, line=144 (Interpreted frame) - java.lang.ref.ReferenceQueue.remove() @bci=2, line=165 (Interpreted frame) - java.lang.ref.Finalizer$FinalizerThread.run() @bci=36, line=216 (Interpreted frame) Thread 20148: (state = BLOCKED) - java.lang.Object.wait(long) @bci=0 (Interpreted frame) - java.lang.Object.wait() @bci=2, line=502 (Interpreted frame) - java.lang.ref.Reference.tryHandlePending(boolean) @bci=54, line=191 (Interpreted frame) - java.lang.ref.Reference$ReferenceHandler.run() @bci=1, line=153 (Interpreted frame) Thread 20137: (state = BLOCKED) - java.lang.Object.wait(long) @bci=0 (Interpreted frame) - org.apache.kafka.clients.Metadata.awaitUpdate(int, long) @bci=63, line=261 (Interpreted frame) - org.apache.kafka.clients.producer.KafkaProducer.waitOnMetadata(java.lang.String, java.lang.Integer, long) @bci=160, line=983 (Interpreted frame) - org.apache.kafka.clients.producer.KafkaProducer.doSend(org.apache.kafka.clients.producer.ProducerRecord, org.apache.kafka.clients.producer.Callback) @bci=19, line=860 (Interpreted frame) - org.apache.kafka.clients.producer.KafkaProducer.send(org.apache.kafka.clients.producer.ProducerRecord, org.apache.kafka.clients.producer.Callback) @bci=12, line=840 (Interpreted frame) - org.apache.kafka.clients.producer.KafkaProducer.send(org.apache.kafka.clients.producer.ProducerRecord) @bci=3, line=727 (Interpreted frame) - org.apache.kafka.log4jappender.KafkaLog4jAppender.append(org.apache.log4j.spi.LoggingEvent) @bci=69, line=283 (Interpreted frame) - org.apache.log4j.AppenderSkeleton.doAppend(org.apache.log4j.spi.LoggingEvent) @bci=106, line=251 (Interpreted frame) - org.apache.log4j.helpers.AppenderAttachableImpl.appendLoopOnAppenders(org.apache.log4j.spi.LoggingEvent) @bci=41, line=66 (Interpreted frame) - org.apache.log4j.Category.callAppenders(org.apache.log4j.spi.LoggingEvent) @bci=26, line=206 (Interpreted frame) - org.apache.log4j.Category.forcedLog(java.lang.String, org.apache.log4j.Priority, java.lang.Object, java.lang.Throwable) @bci=14, line=391 (Interpreted frame) - org.apache.log4j.Category.log(java.lang.String, org.apache.log4j.Priority, java.lang.Object, java.lang.Throwable) @bci=34, line=856 (Interpreted frame) - org.slf4j.impl.Log4jLoggerAdapter.info(java.lang.String) @bci=12, line=305 (Interpreted frame) - org.apache.spark.internal.Logging$class.logInfo(org.apache.spark.internal.Logging, scala.Function0) @bci=29, line=54 (Interpreted frame) - org.apache.spark.SecurityManager.logInfo(scala.Function0) @bci=2, line=44 (Interpreted frame) - org.apache.spark.SecurityManager.setViewAcls(scala.collection.immutable.Set, java.lang.String) @bci=36, line=139 (Interpreted frame) - org.apache.spark.SecurityManager.<init>(org.apache.spark.SparkConf, scala.Option) @bci=158, line=81 (Interpreted frame) - org.apache.spark.deploy.yarn.ApplicationMaster.<init>(org.apache.spark.deploy.yarn.ApplicationMasterArguments) @bci=85, line=70 (Interpreted frame) - org.apache.spark.deploy.yarn.ApplicationMaster$.main(java.lang.String[]) @bci=25, line=802 (Interpreted frame) - org.apache.spark.deploy.yarn.ApplicationMaster.main(java.lang.String[]) @bci=4 (Interpreted frame)
到这里，已经确定是死锁，导致driver一开始就运行停滞，那么当然无法提交executor执行。
具体的死锁稍后分析，先考虑如何解决。从感性认识看，似乎只要不让kafka-client的日志也输出到kafka即可。实验后，发现果然如此：如果只输出org.apache.spark的日志就可以正常执行。
根因分析
从stack的结果看，造成死锁的是如下两个线程：
- kafka-client内部的网络线程spark
- 主入口线程
两个线程其实都是卡在打日志上了，观察堆栈可以发现，两个线程同时持有了同一个log对象。而这个log对象实际上是kafka-appender。而kafka-appender本质上持有kafka-client，及其内部的Metadata对象。log4j的doAppend为了保证线程安全也用synchronized修饰了：
public
synchronized
void doAppend(LoggingEvent event) {
if(closed) {
LogLog.error("Attempted to append to closed appender named ["+name+"].");
return;
}
if(!isAsSevereAsThreshold(event.level)) {
return;
}
Filter f = this.headFilter;
FILTER_LOOP:
while(f != null) {
switch(f.decide(event)) {
case Filter.DENY: return;
case Filter.ACCEPT: break FILTER_LOOP;
case Filter.NEUTRAL: f = f.next;
}
}
this.append(event);
}
于是事情开始了：
- main线程尝试打日志，首先进入了synchronized的doAppend，即获取了
kafka-appender的锁 kafka-appender内部需要调用kafka-client发送日志到kafka，最终调用到Thread 20137展示的，运行到Metadata.awaitUpdate（也是个synchronized方法），内部的wait会尝试获取metadata的锁。（详见https://github.com/apache/kaf...）- 但此时，kafka-producer-network-thread线程刚好进入了上文提到的打
Cluster ID这个日志的这个阶段（update方法也是synchronized的），也就是说kafka-producer-network-thread线程获得了metadata对象的锁 - kafka-producer-network-thread线程要打印日志同样执行synchronized的doAppend，即获取了
kafka-appender的锁
上图main-thread持有了log对象锁，要求获取metadata对象锁；kafka-producer-network-thread持有了metadata对象锁，要求获取log对象锁于是造成了死锁。
总结
DDR爱好者之家 Design By 杰米
广告合作：本站广告合作请联系QQ：858582 申请时备注：广告合作（否则不回）
免责声明：本站资源来自互联网收集,仅供用于学习和交流,请遵循相关法律法规,本站一切资源不代表本站立场,如有侵权、后门、不妥请联系本站删除！
免责声明：本站资源来自互联网收集,仅供用于学习和交流,请遵循相关法律法规,本站一切资源不代表本站立场,如有侵权、后门、不妥请联系本站删除！
DDR爱好者之家 Design By 杰米
暂无评论...
RTX 5090要首发 性能要翻倍！三星展示GDDR7显存
三星在GTC上展示了专为下一代游戏GPU设计的GDDR7内存。
首次推出的GDDR7内存模块密度为16GB，每个模块容量为2GB。其速度预设为32 Gbps（PAM3），但也可以降至28 Gbps，以提高产量和初始阶段的整体性能和成本效益。
据三星表示，GDDR7内存的能效将提高20%，同时工作电压仅为1.1V，低于标准的1.2V。通过采用更新的封装材料和优化的电路设计，使得在高速运行时的发热量降低，GDDR7的热阻比GDDR6降低了70%。
更新日志
2024年11月23日
2024年11月23日
- 凤飞飞《我们的主题曲》飞跃制作[正版原抓WAV+CUE]
- 刘嘉亮《亮情歌2》[WAV+CUE][1G]
- 红馆40·谭咏麟《歌者恋歌浓情30年演唱会》3CD[低速原抓WAV+CUE][1.8G]
- 刘纬武《睡眠宝宝竖琴童谣 吉卜力工作室 白噪音安抚》[320K/MP3][193.25MB]
- 【轻音乐】曼托凡尼乐团《精选辑》2CD.1998[FLAC+CUE整轨]
- 邝美云《心中有爱》1989年香港DMIJP版1MTO东芝首版[WAV+CUE]
- 群星《情叹-发烧女声DSD》天籁女声发烧碟[WAV+CUE]
- 刘纬武《睡眠宝宝竖琴童谣 吉卜力工作室 白噪音安抚》[FLAC/分轨][748.03MB]
- 理想混蛋《Origin Sessions》[320K/MP3][37.47MB]
- 公馆青少年《我其实一点都不酷》[320K/MP3][78.78MB]
- 群星《情叹-发烧男声DSD》最值得珍藏的完美男声[WAV+CUE]
- 群星《国韵飘香·贵妃醉酒HQCD黑胶王》2CD[WAV]
- 卫兰《DAUGHTER》【低速原抓WAV+CUE】
- 公馆青少年《我其实一点都不酷》[FLAC/分轨][398.22MB]
- ZWEI《迟暮的花 (Explicit)》[320K/MP3][57.16MB]