In other words, when a field F depends on a non-stored field G, it also depends
on G's dependencies. This guarantees that whenever a dependency of G is
modified, F will be invalidated and marked to recompute (if necessary).
The transitive closure of dependencies is not computed over stored fields.
Anyway stored fields already trigger the recomputation of their dependent
fields during their recomputation. The performance impact on the loading of a
registry is negligible (less than 1%), and the increase of recomputation
triggers is small (less than 10%).
We sometimes use in Odoo a One2many field with an inverse Integer
field instead of a usual Many2one.
This allow for example in several instances to have a "Many2one" which
can be reference from several models, eg:
Model Ranking:
name = String field
res_id = Integer field
res_model = String field
Model Toy:
rank = One2many [inverse: Ranking -> res_id]
[domain: res_model == Toy]
Model Tool:
rank = One2many [inverse: Ranking -> res_id]
[domain: res_model == Tool]
This enable us to have a shared feature between otherwise unrelated models.
But there was several issue when searching on these One2many:
1) if the Integer Many2one was not stored (eg. it came from an inherits) on
the searched model, this could lead to an error.
2) when we searched:
- by IDs (rank in ['55']) with at least one id not respecting the domain
- by IDs with a negative operator
- with a negative operator on unfound string (rank != "no rank has this")
- with a false value (rank = False)
we would not apply the One2many domain (eg. res_model == Toy) and thus
possibly getting Toy 3 errenously because a Tool 3 was found without
the domain being applied.
This fix modify the search on One2many and for:
1. if the inverse is an Integer not stored field instead of Many2one
manage it.
2. if the field is an Integer field instead of Many2one and there is a
domain on the One2many: apply the domain on the inverse model found.
So only some search on One2many with a domain whose inverse field is an
Integer could be impacted.
This would also be nice to have for all One2many with a domain but the
probability of it being useful versus risk for performance is not judged
high enough.
opw-710508
This removal has led to a necessary refactoring:
- make the setup of field attributes extensible;
- make the instantiation of custom fields extensible;
- delegate model and field reflection to `ir.model` and `ir.model.fields`;
- move the implementation of sparse and serialized fields to module `base_sparse_field`;
Raphael Collet