Apache Beam Python SDK-会话窗口间隔不准确

我正在尝试使用ApacheBeamPython SDK以60分钟的会话间隔处理数据。但实际会话间隔不准确，例如运行应用程序时

3:00:00

或

1:01:00

或

1:50:00

您能帮我找到一个解决方案来解决这个问题并用60分钟的会话处理数据吗

我建立了我的管道作为贝娄

    with Pipeline(options=pipeline_options) as pipeline:
        (
            pipeline
            | "Read" >> ReadFromText(known_args.input, skip_header_lines=1)
            | "Convert" >> ParDo(Convert())
            | "Add Timestamp" >> Map(lambda x: window.TimestampedValue(x, get_timestamp_from_element(x).timestamp()))
            | "Use User ID As Key" >> Map(lambda x: (x["user_id"], x))
            | "Apply Session Window" >> WindowInto(window.Sessions(known_args.session_interval))
            | "Group" >> GroupByKey()
            | "Write To CSV" >> ParDo(WriteToCSV(known_args.output))
        )
        result = pipeline.run()
        result.wait_until_finish()

会话间隔（60分钟）如下所示

    parser.add_argument(
        "--session_interval",
        help="Interval of each session",
        default=60*60) # 60 mins

WriteToCSV
功能处理每个会话的数据。我记录了会话持续时间，但不准确

class WriteToCSV(DoFn):
    def __init__(self, output_path):
        self.output_path = output_path

    def process(self, element, window=DoFn.WindowParam):
        window_start = window.start.to_utc_datetime()
        window_end = window.end.to_utc_datetime()
        duration = window_end - window_start
        logging.info(">>> new %s record(s) in %s session (start %s end %s)", len(click_records), duration, window_start, window_end)
        ....

然后，当我使用DirectRunner在本地运行此应用程序时，我收到了此日志消息

new 5 records in 3:00:00 session (start 2018-10-19 02:00:00 end 2018-10-19 05:00:00)
new 2 records in 1:01:00 session (start 2018-10-19 02:02:00 end 2018-10-19 03:03:00)
new 2 records in 1:50:00 session (start 2018-10-19 03:10:00 end 2018-10-19 05:00:00)

我还将管道部署到数据流，然后得到了相同的结果

new 2 record(s) in 1:50:00 session (start 2018-10-19 11:10:00 end 2018-10-19 13:00:00)
new 2 record(s) in 1:01:00 session (start 2018-10-19 10:02:00 end 2018-10-19 11:03:00)
new 5 record(s) in 3:00:00 session (start 2018-10-19 10:00:00 end 2018-10-19 13:00:00)

在beam管道中，

window.Sessions`中的变量“`known_args.session_interval

”定义了间隙持续时间，即如果某个特定键没有进一步的事件发生，则窗口关闭的持续时间。根据管道为给定密钥处理的事件数，每个会话可以具有不同的开始和结束持续时间。这是用图画来解释的

比如说

Key 1 - 10:00 AM ----|
Key 1 - 10:45 AM     |
Key 1 - 11:30 AM     |====> One Session Window for Key 1 of Duration 4hours 30 minutes
Key 1 - 12:15 PM     |
Key 1 - 01:00 PM ----|
Key 1 - 02:30 PM =========> Start of new session window for Key 1

Key 2 - 10:00 AM-----|
Key 2 - 10:30 AM     |====> One Session window for key 2 of Duration 1:00 hour
Key 2 - 11:00 PM-----|
Key 2 - 12:30 PM =========> Start of new session window for Key 2

如果您对每60分钟分组和处理一次事件感兴趣，那么您需要使用FixedWindows