Paul, ”The end of ETL as we know it" doesn’t make any sense.

In general, nothing new, I’ll say. Instead of doing ETL in “pull” mode where the data from OLTP systems goes to the OLAP systems for performing analytics, you suggest that OLTP systems publish the data in a “publish-subscribe” way, the architectural version of the observer pattern. What is new?

If you are in a “Lamba architecture” for example, where your data exchange is performed with Kafka queues, you can take advantage of the analytics as you described. In this way, you can achieve some benefits, such as a precise schema for your data. But it is a very particular situation.

Why is it a particular situation? Because most of the time, you cannot benefit from this luxury. Most of the time, data to put into the data warehouses or systems for analytics come from operational systems. The relative software simply doesn’t publish any data in a publish-subscribe way.

Possible alternative: to customize all the software that writes some data on OLTP systems to operate with publish-subscribe queues. Bullshits. No advantages, higher costs, less cohesion, sparse ‘logic’ of the application.