[Bug]: Window trigger has lag when using Kafka and GroupByKey on Dataflow Runner

I wrote my stream processing task with Python and Apache Beam. My pipeline has ReadFromKafka, WindowInto fixed windows and GroupByKey. I run it on Dataflow runner. Everything works fine except window triggering which has delay. There is a lag even when I use early AfterProcessing:

```
| 'Read from Kafka' >> ReadFromKafka(consumer_config=consumer_config, topics=[topic])
| 'Convert' >> ParDo(Convert())
| 'Fixed window' >> WindowInto(window.FixedWindows(1)
, trigger=trigger.AfterWatermark(early=trigger.AfterProcessingTime(delay=1))
, accumulation_mode=trigger.AccumulationMode.ACCUMULATING)
| "Group by key" >> GroupByKey()
| "Debug windows" >> ParDo(DebugWindow()).with_output_types(typing.Tuple[bytes, bytes])
```

In the DebugWindow I have added a log to estimate the delay in this way:

```
window_end = window.end.to_utc_datetime()
current_time = datetime.utcnow()
delay = (current_time - window_end).total_seconds()
logging.info(f'window delay: {str(delay)}')
```
The behavior of this delay is strange: it is decreasing and the suddenly jump and increase again. The following image show this behavior:


The left column is the dataflow log time which is trigger time I guess and the right column is the logged delay. As you see there is more than a 1s delay in trigger time (42s to 44s) and the processing delay suddenly jumps to 172s.
What is the problem? I need to have a near real-time (less than 2s) pipeline.

P.S. I have tested increasing number of workers and increasing the number of Kafka partitions but none of them help.

### Issue Priority

Priority: 1 (data loss / total loss of function)

### Issue Components

- [X] Component: Python SDK
- [ ] Component: Java SDK
- [ ] Component: Go SDK
- [ ] Component: Typescript SDK
- [X] Component: IO connector
- [ ] Component: Beam examples
- [ ] Component: Beam playground
- [ ] Component: Beam katas
- [ ] Component: Website
- [ ] Component: Spark Runner
- [ ] Component: Flink Runner
- [ ] Component: Samza Runner
- [ ] Component: Twister2 Runner
- [ ] Component: Hazelcast Jet Runner
- [X] Component: Google Cloud Dataflow Runner