| commit | b2ff4c4f7ec21d41cb173b413bd5aa5feefd7eee | [log] [tgz] |
|---|---|---|
| author | Emil Ejbyfeldt <eejbyfeldt@liveintent.com> | Fri Apr 07 10:14:07 2023 +0900 |
| committer | Hyukjin Kwon <gurwls223@apache.org> | Fri Apr 07 10:14:14 2023 +0900 |
| tree | c9e01bfc757d27cef9033a65d8bf78dc938cdc66 | |
| parent | 1d974a7a78fa2a4d688d5e8606dcd084ab08b220 [diff] |
[SPARK-39696][CORE] Fix data race in access to TaskMetrics.externalAccums ### What changes were proposed in this pull request? This PR fixes a data race around concurrent access to `TaskMetrics.externalAccums`. The race occurs between the `executor-heartbeater` thread and the thread executing the task. This data race is not known to cause issues on 2.12 but in 2.13 ~due this change https://github.com/scala/scala/pull/9258~ (LuciferYang bisected this to first cause failures in scala 2.13.7 one possible reason could be https://github.com/scala/scala/pull/9786) leads to an uncaught exception in the `executor-heartbeater` thread, which means that the executor will eventually be terminated due to missing hearbeats. This fix of using of using `CopyOnWriteArrayList` is cherry picked from https://github.com/apache/spark/pull/37206 where is was suggested as a fix by LuciferYang since `TaskMetrics.externalAccums` is also accessed from outside the class `TaskMetrics`. The old PR was closed because at that point there was no clear understanding of the race condition. JoshRosen commented here https://github.com/apache/spark/pull/37206#issuecomment-1189930626 saying that there should be no such race based on because all accumulators should be deserialized as part of task deserialization here: https://github.com/apache/spark/blob/0cc96f76d8a4858aee09e1fa32658da3ae76d384/core/src/main/scala/org/apache/spark/executor/Executor.scala#L507-L508 and therefore no writes should occur while the hearbeat thread will read the accumulators. But my understanding is that is incorrect as accumulators will also be deserialized as part of the taskBinary here: https://github.com/apache/spark/blob/169f828b1efe10d7f21e4b71a77f68cdd1d706d6/core/src/main/scala/org/apache/spark/scheduler/ShuffleMapTask.scala#L87-L88 which will happen while the heartbeater thread is potentially reading the accumulators. This can both due to user code using accumulators (see the new test case) but also when using the Dataframe/Dataset API as sql metrics will also be `externalAccums`. One way metrics will be sent as part of the taskBinary is when the dep is a `ShuffleDependency`: https://github.com/apache/spark/blob/fbbcf9434ac070dd4ced4fb9efe32899c6db12a9/core/src/main/scala/org/apache/spark/Dependency.scala#L85 with a ShuffleWriteProcessor that comes from https://github.com/apache/spark/blob/fbbcf9434ac070dd4ced4fb9efe32899c6db12a9/sql/core/src/main/scala/org/apache/spark/sql/execution/exchange/ShuffleExchangeExec.scala#L411-L422 ### Why are the changes needed? The current code has a data race. ### Does this PR introduce _any_ user-facing change? It will fix an uncaught exception in the `executor-hearbeater` thread when using scala 2.13. ### How was this patch tested? This patch adds a new test case, that before the fix was applied consistently produces the uncaught exception in the heartbeater thread when using scala 2.13. Closes #40663 from eejbyfeldt/SPARK-39696. Lead-authored-by: Emil Ejbyfeldt <eejbyfeldt@liveintent.com> Co-authored-by: yangjie01 <yangjie01@baidu.com> Signed-off-by: Hyukjin Kwon <gurwls223@apache.org> (cherry picked from commit 6ce0822f76e11447487d5f6b3cce94a894f2ceef) Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
Spark is a unified analytics engine for large-scale data processing. It provides high-level APIs in Scala, Java, Python, and R, and an optimized engine that supports general computation graphs for data analysis. It also supports a rich set of higher-level tools including Spark SQL for SQL and DataFrames, pandas API on Spark for pandas workloads, MLlib for machine learning, GraphX for graph processing, and Structured Streaming for stream processing.
You can find the latest Spark documentation, including a programming guide, on the project web page. This README file only contains basic setup instructions.
Spark is built using Apache Maven. To build Spark and its example programs, run:
./build/mvn -DskipTests clean package
(You do not need to do this if you downloaded a pre-built package.)
More detailed documentation is available from the project site, at “Building Spark”.
For general development tips, including info on developing Spark using an IDE, see “Useful Developer Tools”.
The easiest way to start using Spark is through the Scala shell:
./bin/spark-shell
Try the following command, which should return 1,000,000,000:
scala> spark.range(1000 * 1000 * 1000).count()
Alternatively, if you prefer Python, you can use the Python shell:
./bin/pyspark
And run the following command, which should also return 1,000,000,000:
>>> spark.range(1000 * 1000 * 1000).count()
Spark also comes with several sample programs in the examples directory. To run one of them, use ./bin/run-example <class> [params]. For example:
./bin/run-example SparkPi
will run the Pi example locally.
You can set the MASTER environment variable when running examples to submit examples to a cluster. This can be a mesos:// or spark:// URL, “yarn” to run on YARN, and “local” to run locally with one thread, or “local[N]” to run locally with N threads. You can also use an abbreviated class name if the class is in the examples package. For instance:
MASTER=spark://host:7077 ./bin/run-example SparkPi
Many of the example programs print usage help if no params are given.
Testing first requires building Spark. Once Spark is built, tests can be run using:
./dev/run-tests
Please see the guidance on how to run tests for a module, or individual tests.
There is also a Kubernetes integration test, see resource-managers/kubernetes/integration-tests/README.md
Spark uses the Hadoop core library to talk to HDFS and other Hadoop-supported storage systems. Because the protocols have changed in different versions of Hadoop, you must build Spark against the same version that your cluster runs.
Please refer to the build documentation at “Specifying the Hadoop Version and Enabling YARN” for detailed guidance on building for a particular distribution of Hadoop, including building for particular Hive and Hive Thriftserver distributions.
Please refer to the Configuration Guide in the online documentation for an overview on how to configure Spark.
Please review the Contribution to Spark guide for information on how to get started contributing to the project.
