如何在Google Cloud Dataflow / Apache Beam中并行运行多个WriteToBigQuery？

Question

我想从给定数据的多个事件中分离事件

{"type": "A", "k1": "v1"}
{"type": "B", "k2": "v2"}
{"type": "C", "k3": "v3"}

我想分隔type: A bigquery中的type: A事件到表A ， type:B事件到表B ， type: C事件到表C

这是我通过apache beam python sdk实现的代码，并将数据写入bigquery ，

A_schema = 'type:string, k1:string'
B_schema = 'type:string, k2:string'
C_schema = 'type:string, k2:string'

class ParseJsonDoFn(beam.DoFn):
    A_TYPE = 'tag_A'
    B_TYPE = 'tag_B'
    C_TYPE = 'tag_C'
    def process(self, element):
        text_line = element.trip()
        data = json.loads(text_line)

        if data['type'] == 'A':
            yield pvalue.TaggedOutput(self.A_TYPE, data)
        elif data['type'] == 'B':
            yield pvalue.TaggedOutput(self.B_TYPE, data)
        elif data['type'] == 'C':
            yield pvalue.TaggedOutput(self.C_TYPE, data)

def run():
    parser = argparse.ArgumentParser()
    parser.add_argument('--input',
                      dest='input',
                      default='data/path/data',
                      help='Input file to process.')
    known_args, pipeline_args = parser.parse_known_args(argv)
    pipeline_args.extend([
      '--runner=DirectRunner',
      '--project=project-id',
      '--job_name=seperate-bi-events-job',
    ])
    pipeline_options = PipelineOptions(pipeline_args)
    pipeline_options.view_as(SetupOptions).save_main_session = True
    with beam.Pipeline(options=pipeline_options) as p:
        lines = p | ReadFromText(known_args.input)

    multiple_lines = (
        lines
        | 'ParseJSON' >> (beam.ParDo(ParseJsonDoFn()).with_outputs(
                                      ParseJsonDoFn.A_TYPE,
                                      ParseJsonDoFn.B_TYPE,
                                      ParseJsonDoFn.C_TYPE)))

    a_line = multiple_lines.tag_A
    b_line = multiple_lines.tag_B
    c_line = multiple_lines.tag_C

    (a_line
        | "output_a" >> beam.io.WriteToBigQuery(
                                          'temp.a',
                                          schema = A_schema,
                                          write_disposition = beam.io.BigQueryDisposition.WRITE_TRUNCATE,
                                          create_disposition = beam.io.BigQueryDisposition.CREATE_IF_NEEDED
                                        ))

    (b_line
        | "output_b" >> beam.io.WriteToBigQuery(
                                          'temp.b',
                                          schema = B_schema,
                                          write_disposition = beam.io.BigQueryDisposition.WRITE_TRUNCATE,
                                          create_disposition = beam.io.BigQueryDisposition.CREATE_IF_NEEDED
                                        ))

    (c_line
        | "output_c" >> beam.io.WriteToBigQuery(
                                          'temp.c',
                                          schema = (C_schema),
                                          write_disposition = beam.io.BigQueryDisposition.WRITE_TRUNCATE,
                                          create_disposition = beam.io.BigQueryDisposition.CREATE_IF_NEEDED
                                        ))

    p.run().wait_until_finish()

输出：

INFO:root:start <DoOperation output_banner/WriteToBigQuery output_tags=['out']>
INFO:oauth2client.transport:Attempting refresh to obtain initial access_token
INFO:oauth2client.client:Refreshing access_token
WARNING:root:Sleeping for 150 seconds before the write as BigQuery inserts can be routed to deleted table for 2 mins after the delete and create.
INFO:root:start <DoOperation output_banner/WriteToBigQuery output_tags=['out']>
INFO:oauth2client.transport:Attempting refresh to obtain initial access_token
INFO:oauth2client.client:Refreshing access_token
WARNING:root:Sleeping for 150 seconds before the write as BigQuery inserts can be routed to deleted table for 2 mins after the delete and create.
INFO:root:start <DoOperation output_banner/WriteToBigQuery output_tags=['out']>
INFO:oauth2client.transport:Attempting refresh to obtain initial access_token
INFO:oauth2client.client:Refreshing access_token
WARNING:root:Sleeping for 150 seconds before the write as BigQuery inserts can be routed to deleted table for 2 mins after the delete and create.

但是，这里有两个问题

• bigquery没有数据？
• 从日志看来，代码不是并行运行，而不是顺序运行3次？

我的代码有什么问题吗？或者我缺少什么？

Answer 1

bigquery中没有数据？

当将数据写入BigQuery时，您的代码似乎非常好（ C_schema应该是k3而不是k2 ）。 请记住，您正在流式传输数据，因此，如果在提交流式缓冲区中的数据之前单击“ Preview表”按钮，您将看不到它。 运行SELECT *查询将显示预期结果：

从日志看来，代码不是并行运行，而不是顺序运行3次？

好的，这很有趣。 通过在代码中跟踪WARNING消息，我们可以阅读以下内容：

# if write_disposition == BigQueryDisposition.WRITE_TRUNCATE we delete
# the table before this point.
if write_disposition == BigQueryDisposition.WRITE_TRUNCATE:
  # BigQuery can route data to the old table for 2 mins max so wait
  # that much time before creating the table and writing it
  logging.warning('Sleeping for 150 seconds before the write as ' +
                  'BigQuery inserts can be routed to deleted table ' +
                  'for 2 mins after the delete and create.')
  # TODO(BEAM-2673): Remove this sleep by migrating to load api
  time.sleep(150)
  return created_table
else:
  return created_table

在阅读BEAM-2673和BEAM-2801之后 ，看来这是由于BigQuery接收器存在问题，该问题与DirectRunner一起使用了Streaming API。 重新创建表时，这将导致进程休眠150 s，并且不会并行执行。

相反，如果我们在Dataflow上运行它（使用DataflowRunner ，提供分段和临时存储区路径，以及从GCS加载输入数据），那么它将并行运行三个导入作业。 在下图中看到，所有这三个开始于22:19:45并结束于22:19:56 ：