Improve Runtime Prediciton

[stefan-grafberger]

Description

• Now that the analyses and optimizations are implemented, there are some cases where the runtime prediction is inaccurate:
  • For data corruption and cleaning operators, the runtime predictions are missing, so they will be calculated as 0
  • ~For filter removal, the runtime of children nodes is not increased according to the selectivity of the removed filter~
  • ~Most importantly, the runtime of our current ModelPatches (cleanlab and shapley filtering and retraining) is not updated to reflect the repeated training. Because the models are already fairly expensive, this can result in highly inaccurate predictions for the DataCleaning what-if analysis~
• We want to address, at least, some of these issues to improve the quality of our estimates

[stefan-grafberger]

• For filter addiiton we want to update the optimizer info as well, if we somehow can get a selectivity estimate

[stefan-grafberger]

Some of this was implemented as part of PR #24

[stefan-grafberger]

Some estimates still seem to be particularly bad, see test_data_cleaning_compas for example.

[stefan-grafberger]

We can investigate if this is due to model patching or something different by replacing the data cleaning model patching with a patch that does nothing.

[stefan-grafberger]

Or maybe we can check during execution where the significant mismatch is by adding some code to the dag_excutor