Reshuffle causes all work to be done by one machine

I've got a pipeline that looks like this:


I start the job with 20 workers, and it's able to read from GCS pretty quickly. However, the `Reshuffle` steps seems to all get sent to a single worker ( assuming this because Dataflow scales down to a single worker while `Reshuffle` is doing it's thing ).

I suspect it's because of the `GroupByKey`:


Which seems to be processing all of the elements at once.

Is there a way to prevent this so that this step doesn't cause Dataflow to scale down a bunch of workers and then scale back up once it's done? Maybe assign a random time between `time.Now()` and `time.Now() + 1 hour`, window into 1 minute buckets, then put everything back in the global window before returning the PCollection?

Or is there something else I should be doing later in the pipeline to ensure that `Reshuffle` doesn't all get done on a single worker?

The payload names no repository files or tests. Start by reproducing the Go pipeline on Dataflow and inspecting the Reshuffle and GroupByKey stages; done means establishing whether the single-worker behavior is expected and documenting or implementing a supported way to avoid the scale-down.

Written by the indexing model from the issue text.