編集

次の方法で共有


RxSpark

revoscalepy.RxSpark(hdfs_share_dir: str = '/user/RevoShare\766RR78ROCWFDMK$',
    share_dir: str = '/var/RevoShare\766RR78ROCWFDMK$',
    user: str = '766RR78ROCWFDMK$', name_node: str = None,
    master: str = 'yarn', port: int = None,
    auto_cleanup: bool = True, console_output: bool = False,
    packages_to_load: list = None, idle_timeout: int = 3600,
    num_executors: int = None, executor_cores: int = None,
    executor_mem: str = None, driver_mem: str = None,
    executor_overhead_mem: str = None, extra_spark_config: str = '',
    spark_reduce_method: str = 'auto', tmp_fs_work_dir: str = None,
    persistent_run: bool = False, wait: bool = True, **kwargs)

Description

Creates the compute context for running revoscalepy analysis on Spark. Note that the use of rx_spark_connect() is recommended over RxSpark() as rx_spark_connect() supports persistent mode with in-memory caching. Run help(revoscalepy.rx_spark_connect) for more information.