DS5: How to assure scalability of AI systems?

Note: differentiate between the DS and A perspective

• modularize code and decouple different development steps

  • it is more about the building blocks than it is about the overall design

• define well the applicability domain

  • scaling up does not mean you apply the same tool everywhere w/o any adaptation

• create tests to validate any change that might occur in the system

  • every time you create sth, think how it can break or be misused and create a simple test to make sure it continues to do what it was built to do

• use data management tools that allow scaling

  • try not to lock-in your development with tools that are inflexible

• optimize the performance of you code

  • from parallelization to just vectorising the for-loops blocks in your code

• document your code well

• use software management tools (e.g. reporting, monitoring, etc)

• be mindful of the cost/benefit ratio in training the model vs. making it better