I tend to split things up based on the resources required. We had a decent sized data pipeline that was entirely microservice and serverless and it was a real

I tend to split things up based on the resources required. We had a decent sized data pipeline that was entirely microservice and serverless and it was a real joy to work with. - Our NLP code lived in a service that ran on GPUs - our ingest service used a high ram, but relatively simple CPU service to do in-memory joins cheaply and efficiently - we had a bunch of specialized query services that were just direct requests to AWS services, or a light lambda wrapper around a call to a service. Coordinated things with airflow and it was very easy to maintain and scaling was pretty efficient since we just scaled the pieces that needed it without wasting money on unneeded compute.