java - 您如何使用 Google DataProc Java 客户端使用关联 GS 存储桶中的 jar 文件和类提交 Spark 作业？

Question

我需要触发 Spark 作业以使用 API 调用从 JSON 文件聚合数据。我使用 spring-boot 来创建资源。因此，解决方案的步骤如下：

1. 用户使用 json 文件作为输入发出 POST 请求
2. JSON 文件存储在与 dataproc 集群关联的 google 存储桶中。
3. 使用指定的 jars、类从 REST 方法中触发聚合 Spark 作业，参数是 json 文件链接。

我希望使用 Dataproc 的 Java 客户端而不是控制台或命令行来触发作业。你怎么做呢？

Answer 1

We're hoping to have a more thorough guide shortly on the official documentation, but to get started, visit the following API overview: https://developers.google.com/api-client-library/java/apis/dataproc/v1

It includes links to the Dataproc javadocs; if your server is making calls on behalf of your own project and not on behalf of your end-users' Google projects, then you probably want the keyfile-based service-account auth explained here to create the Credential object you use to initialize the Dataproc client stub.

As for the dataproc-specific parts, this just means you add the following dependency to your Maven pomfile if using Maven:

<project>
  <dependencies>
    <dependency>
      <groupId>com.google.apis</groupId>
      <artifactId>google-api-services-dataproc</artifactId>
      <version>v1-rev4-1.21.0</version>
    </dependency>
  </dependencies>
</project>

And then you'll have code like:

Dataproc dataproc = new Dataproc.Builder(new NetHttpTransport(), new JacksonFactory(), credential)
    .setApplicationName("my-webabb/1.0")
    .build();
dataproc.projects().regions().jobs().submit(
    projectId, "global", new SubmitJobRequest()
        .setJob(new Job()
            .setPlacement(new JobPlacement()
                .setClusterName("my-spark-cluster"))
            .setSparkJob(new SparkJob()
                .setMainClass("FooSparkJobMain")
                .setJarFileUris(ImmutableList.of("gs://bucket/path/to/your/spark-job.jar"))
                .setArgs(ImmutableList.of(
                    "arg1", "arg2", "arg3")))))
    .execute();

Since different intermediary servers may do low-level retries or your request may throw an IOException where you don't know whether the job-submission succeeded or not, an addition step you may want to take is to generate your own jobId; then you know what jobId to poll on to figure out if it got submitted even if your request times out or throws some unknown exception:

import java.util.UUID;

...

Dataproc dataproc = new Dataproc.Builder(new NetHttpTransport(), new JacksonFactory(), credential)
    .setApplicationName("my-webabb/1.0")
    .build();

String curJobId = "json-agg-job-" + UUID.randomUUID().toString();
Job jobSnapshot = null;
try {
  jobSnapshot = dataproc.projects().regions().jobs().submit(
      projectId, "global", new SubmitJobRequest()
          .setJob(new Job()
              .setReference(new JobReference()
                   .setJobId(curJobId))
              .setPlacement(new JobPlacement()
                  .setClusterName("my-spark-cluster"))
              .setSparkJob(new SparkJob()
                  .setMainClass("FooSparkJobMain")
                  .setJarFileUris(ImmutableList.of("gs://bucket/path/to/your/spark-job.jar"))
                  .setArgs(ImmutableList.of(
                      "arg1", "arg2", "arg3")))))
      .execute();
} catch (IOException ioe) {
  try {
    jobSnapshot = dataproc.projects().regions().jobs().get(
        projectId, "global", curJobId).execute();
    logger.info(ioe, "Despite exception, job was verified submitted");
  } catch (IOException ioe2) {
    // Handle differently; if it's a GoogleJsonResponseException you can inspect the error
    // code, and if it's a 404, then it means the job didn't get submitted; you can add retry
    // logic in that case.
  }
}

// We can poll on dataproc.projects().regions().jobs().get(...) until the job reports being
// completed or failed now.