Spark3.1.2 standalone集群s3a目录提交器OOM无法创建线程问题求助

问题背景 #

使用Spark 3.1.2 standalone集群，启用s3a directory committer后作业稳定性和性能有明显提升，但近期频繁出现Driver OOM错误，报错为无法创建原生线程。

错误日志 #

An error occurred while calling None.org.apache.spark.api.java.JavaSparkContext.

: java.lang.OutOfMemoryError: unable to create native thread: possibly out of memory or process/resource limits reached
    at java.base/java.lang.Thread.start0(Native Method)
    at java.base/java.lang.Thread.start(Thread.java:803)
    at java.base/java.util.concurrent.ThreadPoolExecutor.addWorker(ThreadPoolExecutor.java:937)
    at java.base/java.util.concurrent.ThreadPoolExecutor.execute(ThreadPoolExecutor.java:1343)
    at java.base/java.util.concurrent.AbstractExecutorService.submit(AbstractExecutorService.java:118)
    at java.base/java.util.concurrent.Executors$DelegatedExecutorService.submit(Executors.java:714)
    at org.apache.spark.rpc.netty.DedicatedMessageLoop.$anonfun$new$1(MessageLoop.scala:174)
    at org.apache.spark.rpc.netty.DedicatedMessageLoop.$anonfun$new$1$adapted(MessageLoop.scala:173)
    at scala.collection.immutable.Range.foreach(Range.scala:158)
    at org.apache.spark.rpc.netty.DedicatedMessageLoop.<init>(MessageLoop.scala:173)
    at org.apache.spark.rpc.netty.Dispatcher.liftedTree1$1(Dispatcher.scala:75)
    at org.apache.spark.rpc.netty.Dispatcher.registerRpcEndpoint(Dispatcher.scala:72)
    at org.apache.spark.rpc.netty.NettyRpcEnv.setupEndpoint(NettyRpcEnv.scala:136)
    at org.apache.spark.storage.BlockManager.<init>(BlockManager.scala:231)
    at org.apache.spark.SparkEnv$.create(SparkEnv.scala:394)
    at org.apache.spark.SparkEnv$.createDriverEnv(SparkEnv.scala:189)
    at org.apache.spark.SparkContext.createSparkEnv(SparkContext.scala:277)
    at org.apache.spark.SparkContext.<init>(SparkContext.scala:458)
    at org.apache.spark.api.java.JavaSparkContext.<init>(JavaSparkContext.scala:58)
    at java.base/jdk.internal.reflect.NativeConstructorAccessorImpl.newInstance0(Native Method)
    at java.base/jdk.internal.reflect.NativeConstructorAccessorImpl.newInstance(NativeConstructorAccessorImpl.java:62)
    at java.base/jdk.internal.reflect.DelegatingConstructorAccessorImpl.newInstance(DelegatingConstructorAccessorImpl.java:45)
    at java.base/java.lang.reflect.Constructor.newInstance(Constructor.java:490)
    at py4j.reflection.MethodInvoker.invoke(MethodInvoker.java:247)
    at py4j.reflection.ReflectionEngine.invoke(ReflectionEngine.java:357)
    at py4j.Gateway.invoke(Gateway.java:238)
    at py4j.commands.ConstructorCommand.invokeConstructor(ConstructorCommand.java:80)
    at py4j.commands.ConstructorCommand.execute(ConstructorCommand.java:69)
    at py4j.GatewayConnection.run(GatewayConnection.java:238)
    at java.base/java.lang.Thread.run(Thread.java:834)

排查信息 #

1. 线程栈Dump显示Driver上存在超过5000个处于WAITING状态的s3-committer-pool线程，远超配置的线程数上限，部分线程信息如下：

Thread ID   Thread Name Thread State    Thread Locks
1047    s3-committer-pool-0 WAITING 
1449    s3-committer-pool-0 WAITING 
1468    s3-committer-pool-0 WAITING 
1485    s3-committer-pool-0 WAITING 
1505    s3-committer-pool-0 WAITING 
1524    s3-committer-pool-0 WAITING 
1529    s3-committer-pool-0 WAITING 
1544    s3-committer-pool-0 WAITING 
1549    s3-committer-pool-0 WAITING 
1809    s3-committer-pool-0 WAITING 
1972    s3-committer-pool-0 WAITING 
1998    s3-committer-pool-0 WAITING 
2022    s3-committer-pool-0 WAITING 
2043    s3-committer-pool-0 WAITING 
2416    s3-committer-pool-0 WAITING 
2453    s3-committer-pool-0 WAITING 
2470    s3-committer-pool-0 WAITING 
2517    s3-committer-pool-0 WAITING 
2534    s3-committer-pool-0 WAITING 
2551    s3-committer-pool-0 WAITING 
2580    s3-committer-pool-0 WAITING 
2597    s3-committer-pool-0 WAITING 
2614    s3-committer-pool-0 WAITING 
2631    s3-committer-pool-0 WAITING 
2726    s3-committer-pool-0 WAITING 
2743    s3-committer-pool-0 WAITING 
2763    s3-committer-pool-0 WAITING 
2780    s3-committer-pool-0 WAITING 
2819    s3-committer-pool-0 WAITING 
2841    s3-committer-pool-0 WAITING 
2858    s3-committer-pool-0 WAITING 
2875    s3-committer-pool-0 WAITING 
2925    s3-committer-pool-0 WAITING 
2942    s3-committer-pool-0 WAITING 
2963    s3-committer-pool-0 WAITING 
2980    s3-committer-pool-0 WAITING 
3020    s3-committer-pool-0 WAITING 
3037    s3-committer-pool-0 WAITING 
3055    s3-committer-pool-0 WAITING 
3072    s3-committer-pool-0 WAITING 
3127    s3-committer-pool-0 WAITING 
3144    s3-committer-pool-0 WAITING 
3163    s3-committer-pool-0 WAITING 
3180    s3-committer-pool-0 WAITING 
3222    s3-committer-pool-0 WAITING 
3242    s3-committer-pool-0 WAITING 
3259    s3-committer-pool-0 WAITING 
3278    s3-committer-pool-0 WAITING 
3418    s3-committer-pool-0 WAITING 
3435    s3-committer-pool-0 WAITING 
3452    s3-committer-pool-0 WAITING 
3469    s3-committer-pool-0 WAITING 
3486    s3-committer-pool-0 WAITING 
3491    s3-committer-pool-0 WAITING 
3501    s3-committer-pool-0 WAITING 
3508    s3-committer-pool-0 WAITING 
4029    s3-committer-pool-0 WAITING 
4093    s3-committer-pool-0 WAITING 
4658    s3-committer-pool-0 WAITING 
4666    s3-committer-pool-0 WAITING 
4907    s3-committer-pool-0 WAITING 
5102    s3-committer-pool-0 WAITING 
5119    s3-committer-pool-0 WAITING 
5158    s3-committer-pool-0 WAITING 
5175    s3-committer-pool-0 WAITING 
5192    s3-committer-pool-0 WAITING 
5209    s3-committer-pool-0 WAITING 
5226    s3-committer-pool-0 WAITING 
5395    s3-committer-pool-0 WAITING 
5634    s3-committer-pool-0 WAITING 
5651    s3-committer-pool-0 WAITING 
5668    s3-committer-pool-0 WAITING 
5685    s3-committer-pool-0 WAITING 
5702    s3-committer-pool-0 WAITING 
5722    s3-committer-pool-0 WAITING 
5739    s3-committer-pool-0 WAITING 
6144    s3-committer-pool-0 WAITING 
6167    s3-committer-pool-0 WAITING 
6289    s3-committer-pool-0 WAITING 
6588    s3-committer-pool-0 WAITING 
6628    s3-committer-pool-0 WAITING 
6645    s3-committer-pool-0 WAITING 
6662    s3-committer-pool-0 WAITING 
6675    s3-committer-pool-0 WAITING 
6692    s3-committer-pool-0 WAITING 
6709    s3-committer-pool-0 WAITING 
7049    s3-committer-pool-0 WAITING 

2. 当前s3a相关配置如下，配置的线程数上限远低于实际观测到的线程数：

fs.s3a.threads.max  100 
fs.s3a.connection.maximum  1000 
fs.s3a.committer.threads 16   
fs.s3a.max.total.tasks  5
fs.s3a.committer.name   directory
fs.s3a.fast.upload.buffer                 disk
io.file.buffer.size                                1048576
mapreduce.outputcommitter.factory.scheme.s3a    - org.apache.hadoop.fs.s3a.commit.S3ACommitterFactory

3. 已测试多个版本的spark-hadoop-cloud依赖，问题均可稳定复现。

问题成因 #

1. 核心原因为Hadoop已知bug HADOOP-17260：s3a committer绑定的线程池在S3AFileSystem实例关闭时没有被正确销毁，每初始化一个新的S3AFileSystem实例就会生成一个全新的s3-committer-pool线程池，闲置线程无法回收最终堆积到数千个，超出系统进程资源限制触发OOM。
2. 你当前测试的所有spark-hadoop-cloud版本对应的底层Hadoop依赖均未合入该bug的修复逻辑，因此更换依赖包后问题仍然稳定复现。
3. Spark Driver运行作业过程中会频繁创建新的S3AFileSystem实例访问S3存储，加速了线程泄漏的速度，短时间内就能达到进程线程数上限。

修复方案 #

根治方案 #

升级底层Hadoop依赖到3.3.1及以上版本，该版本已经合入HADOOP-17260的修复逻辑，线程池会随FileSystem实例关闭正常销毁，不会出现线程堆积问题。若使用CDH发行版，升级到7.2.15及以上版本即可解决问题。

临时规避方案 #

如果暂时无法升级Hadoop版本，可以通过添加以下配置强制S3AFileSystem实例全局复用，避免频繁创建新实例触发线程泄漏：

spark.hadoop.fs.s3a.impl.disable.cache false
spark.hadoop.fs.s3a.cache.max.size 20

该配置会让Spark全局复用已创建的S3AFileSystem实例，大幅降低新实例创建频率，缓解线程堆积速度。

临时缓解方案 #

调高Driver进程的ulimit最大线程数限制（将ulimit -u参数调整到10000以上），可以避免作业短时间内崩溃，但无法彻底解决线程泄漏问题，适合作为过渡方案使用。

内容的提问来源于stack exchange，提问作者Ofer Eliassaf