Making KFP Simple Again!
While using KFP as a production back-end @ irobot my team hit a number of pain points for my team. They needed to:
- run locally and be productive
- run on a cluster and not fight tooling, specifically KFP itself.
Here is an approach that I found to be very productive:
- Use light weight component pipeline composition.
- Standardized components (using dockerfiles) to simplify debugging in different contexts.
An example of this type of system can be seen here: