benjello
commented
7 years ago Some comments:
- When computing an individual variable, it might be doable to specify sufficient input variables. But when you look at the whole system (decompositions, fiche de paie, even mes-aides as a whole etc) to identify the different pass-through it is very painful to provide sufficient initialized variables. Defaulting to the least harmful default value seems a reasonable choice. But I can understand that one may want to have a mode where errors pop up when a variable is missing, but the user should have the option to choose the first behavior (I just saw that you rejected that option and I do not need it).
- I agree on the consensus on
base_functionand defaulting to the simulation period. - I greatly suggest that the time-based inference should be provided as an option per variable.
- A backstop date or period of time might also be implemented. We have some variable whose calculation undergoing infinite loops in the pas (X(t) needs Y(t-1) which in turn needs X(t-2)) that could be halted by requiring that passing a certain date/period the default (or custom default) or the last known value is used (We should get rid of the confusing
max_nb_cycles). - Choosing the input setter should be possible. For example I know that salaire_imposable is a monthly variable but I will provide yearly variable (eventually for some past years), I should be able to config the fact that the value should be divided by 12 (and avoiding multiple periods storing).
I hope this helps and keep on the good work !
MattiSG
Morendil
This issue handles a part of #458.
For
base_function, the issue rose with the formularsa, where macrosimulation (where missing values should be the default ones (or raise warnings, @benjello?) and production (where missing values should be inferred) expected behaviours differ.Default values in input variables are the source of the issue: if a formula needs values that are undefined (e.g. because they are in previous months from the calculation), it won't crash nor log anything because the input variables return their default values. This inference behaviour seems to be fine for macrosimulation, but production would expect a time-based inference, where the value is copied from its closest defined value rather than a global default one. Changing the inference mode means changing the
base_function, but that raises conflicts between uses.There is a consensus between @cbenz @michelbl @fpagnoux that
base_functionis incomprehensible and counterintuitive. It is at least misnamed and it should be more explicit that it is currently used purely as an extrapolation mechanism.There is also a consensus that defaulting to the simulation period when requesting to compute a formula is not a proper behaviour: this piece of information should be mandatory to avoid incomprehensible behaviour on omission.
Discussed possible solutions
Candidates
Eliminated
Suggested action plan
base_function, check how many tests fail and evaluate the amount of work to refactor the inputs.