使用SBT编译Spark子项目

前言

最近为了解决Spark2.1的Bug，对Spark的源码做了不少修改，需要对修改的代码做编译测试，如果编译整个Spark项目快的话，也得半小时左右，所以基本上是改了哪个子项目就单独对那个项目编译打包。

Spark官方已经给出了如何使用mvn单独编译子项目的方法：http://spark.apache.org/docs/latest/building-spark.html#building-submodules-individually

使用mvn单独编译子项目是节约了不少时间。但是频繁的改动项目，每次用mvn编译还是挺耗时间的。

之前看官方文档提到，对于开发者，为了提高效率，推荐使用sbt编译。于是，又查了下文档资料：http://spark.apache.org/developer-tools.html

咦，看到：Running Build Targets For Individual Projects，内容如下：

$ # sbt
$ build/sbt package
$ # Maven
$ build/mvn package -DskipTests -pl assembly

这不是坑么，虽然没怎么用sbt编译过Spark，但是sbt俺还是用过的。build/sbt package明明是编译整个项目的好吧，这哪是编译子项目啊。

翻遍官方所有跟编译有关的资料，无果。

最后，研究了下Spark的sbt定义，也就是下project/SparkBuild.scala文件，找到了使用sbt编译子项目的方法。

使用sbt编译子项目

下面是对spark-core重新编译打包的方法，我们需要使用REPL模式，大致的流程如下：

➜  spark git:(branch-2.1.0) ✗ ./build/sbt -Pyarn -Phadoop-2.6 -Phive              
...
[info] Set current project to spark-parent (in build file:/Users/stan/Projects/spark/)
> project core
[info] Set current project to spark-core (in build file:/Users/stan/Projects/spark/)
> package
[info] Updating {file:/Users/stan/Projects/spark/}tags...
[info] Resolving jline#jline;2.12.1 ...
...
[info] Packaging /Users/stan/Projects/spark/core/target/scala-2.11/spark-core_2.11-2.1.0.jar ...
[info] Done packaging.
[success] Total time: 213 s, completed 2017-2-15 16:58:15

最后将spark-core_2.11-2.1.0.jar替换到jars或者assembly/target/scala-2.11/jars目录下就可以了。

选择的子项目的关键是project命令，如何知道有哪些定义好的子项目呢？这个还得参考project/SparkBuild.scala中BuildCommons的定义：

object BuildCommons {

  private val buildLocation = file(".").getAbsoluteFile.getParentFile

  val sqlProjects@Seq(catalyst, sql, hive, hiveThriftServer, sqlKafka010) = Seq(
    "catalyst", "sql", "hive", "hive-thriftserver", "sql-kafka-0-10"
  ).map(ProjectRef(buildLocation, _))

  val streamingProjects@Seq(
    streaming, streamingFlumeSink, streamingFlume, streamingKafka, streamingKafka010
  ) = Seq(
    "streaming", "streaming-flume-sink", "streaming-flume", "streaming-kafka-0-8", "streaming-kafka-0-10"
  ).map(ProjectRef(buildLocation, _))

  val allProjects@Seq(
    core, graphx, mllib, mllibLocal, repl, networkCommon, networkShuffle, launcher, unsafe, tags, sketch, _*
  ) = Seq(
    "core", "graphx", "mllib", "mllib-local", "repl", "network-common", "network-shuffle", "launcher", "unsafe",
    "tags", "sketch"
  ).map(ProjectRef(buildLocation, _)) ++ sqlProjects ++ streamingProjects

  val optionallyEnabledProjects@Seq(mesos, yarn, java8Tests, sparkGangliaLgpl,
    streamingKinesisAsl, dockerIntegrationTests) =
    Seq("mesos", "yarn", "java8-tests", "ganglia-lgpl", "streaming-kinesis-asl",
      "docker-integration-tests").map(ProjectRef(buildLocation, _))

  val assemblyProjects@Seq(networkYarn, streamingFlumeAssembly, streamingKafkaAssembly, streamingKafka010Assembly, streamingKinesisAslAssembly) =
    Seq("network-yarn", "streaming-flume-assembly", "streaming-kafka-0-8-assembly", "streaming-kafka-0-10-assembly", "streaming-kinesis-asl-assembly")
      .map(ProjectRef(buildLocation, _))

  val copyJarsProjects@Seq(assembly, examples) = Seq("assembly", "examples")
    .map(ProjectRef(buildLocation, _))

  val tools = ProjectRef(buildLocation, "tools")
  // Root project.
  val spark = ProjectRef(buildLocation, "spark")
  val sparkHome = buildLocation

  val testTempDir = s"$sparkHome/target/tmp"

  val javacJVMVersion = settingKey[String]("source and target JVM version for javac")
  val scalacJVMVersion = settingKey[String]("source and target JVM version for scalac")
}

我们看下这个例子：

  val sqlProjects@Seq(catalyst, sql, hive, hiveThriftServer, sqlKafka010) = Seq(
    "catalyst", "sql", "hive", "hive-thriftserver", "sql-kafka-0-10"
  ).map(ProjectRef(buildLocation, _))

这是对sql项目定义的子项目，有：catalyst, sql, hive, hiveThriftServer, sqlKafka010。

我们如果需要编译catalyst这个项目，只需要进入sbt：project catalyst选择catalyst项目就可以了，后面使用的compile、package等命令都是针对这个项目的。

结语

这下可以爽爽的编译Spark了。

还有一些有用的编译技巧，去参考http://spark.apache.org/developer-tools.html就可以了。