[Bug]: GroupByKey() emits only key but not values for streaming job on a FixedWindows() with early trigger

[elitongadotti]

What happened?

Setup:

• Python SDK, version 2.60.0
• Dataflow Runner
• Streaming Job

Take a streaming job that reads from a PubSub subscription, do some transformations and sink it to a PubSub Topic.

The message has a lot of duplicate "rows", I need to expose the primary key for an element and GroupByKey() to be able to fetch the latest one. Once I expose pk and group by it, the step downstream might have keys with no values.

If I remove the early=trigger.AfterProcessingTime(some_time) it did work as expected (keys and values are emitted together).

It does not make sense to emit to the downstream steps only keys but not values. Both are part of the same output tuple and they cannot be emitted separately.

Here's part of the code:

  (
      to_sync
      | "AddTimestamp" >> beam.ParDo(AddTimestamp())
      | "Window" >> beam.WindowInto(
          windowfn=window.FixedWindows(60), 
          trigger=trigger.AfterWatermark(early=trigger.AfterProcessingTime(30)),
          accumulation_mode=trigger.AccumulationMode.DISCARDING
      )
      | "GroupAndFetchLatest" >> GroupAndFetchLatest()
      ...
  )
  
 class GroupAndFetchLatest(beam.PTransform):
"""Expose record PK, group by it and fetch the latest record (timestamp)"""
def expand(self, pcoll):
    return (
        pcoll
        | "ExposeKey" >> beam.Map(lambda x: (x["mykey"], x)) 
        | "Group" >> beam.GroupByKey()
        | "GetLatest" >> beam.ParDo(GetLatest())
    )


class GetLatest(beam.DoFn):
    def process(self, grouped):
        key, values = grouped
        if values: # Here, I had to add this "if" to avoid keys without values.
            latest = max(values, key=lambda o: o["__ts__"])
            yield MyObject(
                PK=key,
                row=latest
            )

Issue Priority

Priority: 2 (default / most bugs should be filed as P2)

Issue Components

• Component: Python SDK
• Component: Java SDK
• Component: Go SDK
• Component: Typescript SDK
• Component: IO connector
• Component: Beam YAML
• Component: Beam examples
• Component: Beam playground
• Component: Beam katas
• Component: Website
• Component: Infrastructure
• Component: Spark Runner
• Component: Flink Runner
• Component: Samza Runner
• Component: Twister2 Runner
• Component: Hazelcast Jet Runner
• Component: Google Cloud Dataflow Runner

[elitongadotti]

.set-labels dataflow,python

[github-actions]

Label groupbykey cannot be managed because it does not exist in the repo. Please check your spelling.