嗨，我正在尝试读取apachebeam上的一个Parquet文件，然后应用一个预处理函数，最后将数据集拆分为验证和训练。到目前为止，管道在数据集拆分之前工作。

def beam_transformation(working_dir, eval_percent=20.0):
  '''
  METADATA
  '''
  RAW_DATA_FEATURE_SPEC = dict(
      [(name, tf.io.FixedLenFeature([], tf.string))  for name in CATEGORICAL_FEATURE_KEYS]  +
      [(name, tf.io.VarLenFeature(tf.string))  for name in OTHER_CATEGORICAL_FEATURE_KEYS]  +
      [(name, tf.io.FixedLenFeature([], tf.int64))                 for name in LABEL_KEY]
  )

  RAW_DATA_METADATA = tft.tf_metadata.dataset_metadata.DatasetMetadata(
      tft.tf_metadata.dataset_schema.schema_utils.schema_from_feature_spec(RAW_DATA_FEATURE_SPEC))    

  tf_transform_output = tft.TFTransformOutput(working_dir)

  '''
  Beam TensorFlow Transform Pipeline
  '''
  with beam.Pipeline() as pipeline:
    with tft_beam.Context(temp_dir=working_dir):
      # Read data
      raw_data = pipeline | 'ReadTrainData' >> beam.io.ReadFromParquet(f'{working_dir}/dataset.parquet')

      # Transform data    
      transformed_dataset_and_metadata, transform_fn = (
        raw_data, RAW_DATA_METADATA) | tft_beam.AnalyzeAndTransformDataset(preprocessing_fn)
      transformed_data, transformed_metadata = transformed_dataset_and_metadata

      # Split the dataset into a training set and an evaluation set
      assert 0 < eval_percent < 100, 'eval_percent must in the range (0-100)'
      train_dataset, eval_dataset = (
      transformed_data
      | 'Split dataset' >> beam.Partition(
      lambda elem, _: int(random.uniform(0, 100) < eval_percent), 2))

我以这个网站为例：https://github.com/googlecloudplatform/cloudml-samples/blob/master/molecules/preprocess.py
问题似乎是因为在拆分数据之前必须先收集数据。我收到的错误如下：
异常：似乎您正试图从广播变量、操作或转换引用sparkcontext。sparkcontext只能在驱动程序上使用，不能在工作程序上运行的代码中使用。有关更多信息，请参阅spark-5063。