Ensuring messages are processed and emitted in-order

[Hadi Zhang <ha...@yelp.com>]

We are using the Beam 2.20 Python SDK on a Flink 1.9 runner. Our
messages originate from a custom source that consumes messages from a
Kafka topic and emits them in the order of their Kafka offsets to a
DoFn. After this DoFn processes the messages, they are emitted to a
custom sink that sends messages to a Kafka topic.

We want to process those messages in the order in which we receive
them from Kafka and then emit them to the Kafka sink in the same
order, but based on our understanding Beam does not provide an
in-order transport. However, in practice we noticed that with a Python
SDK worker on Flink and a parallelism setting of 1 and one sdk_worker
instance, messages seem to be both processed and emitted in order. Is
that implementation-specific in-order behavior something that we can
rely on, or is it very likely that this will break at some future
point?

In case it's not recommended to depend on that behavior what is the
best approach for in-order processing?
https://stackoverflow.com/questions/45888719/processing-total-ordering-of-events-by-key-using-apache-beam
recommends to order events in a heap, but according to our
understanding this approach will only work when directly writing to an
external system.