1
The problem with building ML based production systems is that software engineering is more like building a house but ML solutions are more like a ship.
Back to archive
Thread
4 tweets
2
ML solutions aren’t about that one pipeline. You need to retrain, adapt features, and constantly monitor and experiment.
3
Software also needs to be built for change, but in ML, it needs to be done in a way that minimizes manual work because you want to be fast. Changing features shouldn‘t require people touching all parts of the system for a couple of weeks.
4
I have yet to see a solution that does this really well. The fragmented toolset adds to this. Pandas prototypes in notebooks become Spark pipelines and Python microservices and so on.