[jira] [Created] (BEAM-7825) Python's DirectRunner emits multiple panes per window and does not discard late data

["Alexey Strokach (JIRA)" <ji...@apache.org>]

Alexey Strokach created BEAM-7825:
-------------------------------------

             Summary: Python's DirectRunner emits multiple panes per window and does not discard late data
                 Key: BEAM-7825
                 URL: https://issues.apache.org/jira/browse/BEAM-7825
             Project: Beam
          Issue Type: Bug
          Components: sdk-py-core
    Affects Versions: 2.13.0
         Environment: OS: Debian rodete.
Beam versions: 2.15.0.dev.
Python versions: Python 2.7, Python 3.7
            Reporter: Alexey Strokach


The documentation for Beam's Windowing and Triggers functionality [states that|https://beam.apache.org/documentation/programming-guide/#triggers] _"if you use Beam’s default windowing configuration and default trigger, Beam outputs the aggregated result when it estimates all data has arrived, and discards all subsequent data for that window"_. However, it seems that the current behavior of Python's DirectRunner is inconsistent with both of those points. In my experience, DirectRunner will process every data point that it reads from the input stream, irrespective of whether or not the timestamp of that data point is older than the timestamps of the windows that have already been processed. Furthermore, it regularly generates multiple "panes" for the same window, apparently disregarding the notion of a watermark?

An integration test demonstrating the inconsistencies between DirectRunner and Dataflow is provided in the linked PR.

Until the limitations of DirectRunner are addressed, maybe they should be listed on the [DirectRunner documentation page](https://beam.apache.org/documentation/runners/direct/)?

As far as I understand (I have not seen this explicitly documented anywhere), in the case of Cloud Dataflow, the pipeline will first process all elements that have accumulated in a PubSub subscription before the start of the pipeline, and will then process all new elements which have a timestamp within a certain narrow range of the current time (UTC). Would this be the behavior that DirectRunner should be trying to emulate? While Dataflow does emit a single pane per window (by default) and discards late data, it may be too eager in what data it calls "late"?




--
This message was sent by Atlassian JIRA
(v7.6.14#76016)