# Main quest

Refactor the `load_views` API so it no longer sends multiple times the same
fields description.

e.g.
When `load_views` is called to get the kanban, tree and form views,
the list of fields of the model was sent 4 times:
- Once for each view, with only the fields used in the view,
  in `['fields_views']['kanban']['fields']` for instance
- Once globally, with all the fields of the model, in `['fields']`

The goal of this revision is to change that so it sends the list of all fields
only once.

In addition, if a view contains x2many fields,
the fields description of the comodel is also sent.
It was sent in the `views` key of the view fields dict.
e.g.
When calling `load_views` of `res.partner` to get the kanban,
tree and form views,
the `res.partner` fields description was actually sent 6 times:
- Once for each view
- Once globally
- Once for each view of the many2many field `child_ids` of the form view, in 
  - `['fields_views']['form']['fields']['child_ids']['views']['kanban']['fields']`
  - `['fields_views']['form']['fields']['child_ids']['views']['form']['fields']`

The change suggested in this revision is to:
- Remove the fields description for each view in `['fields_views']`.
  As it no longer contains the fields,
  the key becomes `['views']` instead of `['fields_views']`.
- Replace the dict key `['fields']` by `['models']`,
  which is a dict with as key the model name and as values
  the model fields description. It contains the fields description
  for all models implied in the view:
  the model of the main view and the model of all one2many and many2many fields.

With this change, the fields description will only be sent once by model
implied in the view.

In addition, the web client was getting the information about the fields
sometimes in the global fields description list (e.g. `['fields']`),
sometimes in the fields description list of the view type
(e.g. `['fields_views']['form']['fields']`),
making it a pain to try to make changes / performance gain
in these field description dictionaries, because you never knew in which dict
the web client was getting its info.
Now, as there is only one place to get the fields description from,
it's clearer and cleaner.

# Secondary quests

- one2many and many2many fields views are passed directly in the main view
  architecture rather than being put in the `views` key
  of the field description.
  This is actually easier to treat by the web client,
  and this will allow in a future work to cache an entire view in one block
  of text rather than having to combine multiple cached blocks of text
  to return one view.
- one2many and many2many fields which do not have directly embedded views
  have their views directly injected in the architecture,
  so the web client doesn't have to do RPC calls to `load_views`
  for each one2many and many2many fields not having embedded views.
  For instance, this allow to reduce the number of RPC calls to `load_views`
  from 8 to 1 when loading the form of `product.product`.
  Currently, this behavior is limited to 1 level deep but we consider making it
  go all the way down in future works. We did not do it for the moment because
  in certain cases it rises the processing time and the size (bytes) too much.
  e.g. the sale.order view can be 5 levels deep,
  meaning you can reach 4 dialogs on top the main view.
  ```
  sale.order form > order_line > sale.order.line form > invoice_lines > 
  account.move.line form > asset_ids > account.asset form >
  depreciation_move_ids > account.move form.
  ```
  This will also benefit in future works to cache an entire view in one block
  of text rather to having to combine multiple cached block of text
  to get one view.
- `fields_view_get` becomes `get_view`.
  As it no longer returns the fields description,
  keeping the `fields` in the name `fields_view_get` no longer makes sense.
  Hence removing `fields` from the method name, it becomes `view_get`.
  As it gets renamed anyway, we take the opportunity to rename it `get_view`,
  which is more in line with the general getter/setter guidelines
  in the model object world.
- `_fields_view_get` becomes `_get_view`. For the same reasons than above.
- `load_views` becomes `get_views`.
  This is not mandatory, there is no technical reason to rename `load_views` as
  it practically sends the same info as before,
  the view architectures and their fields description. Just in another way.
  We just take the opportunity of this pull request to suggest a cleaner API:
  `_get_view`, `get_view` and `get_views`.
- Arguments `toolbar=False, submenu=False` fo the methods
  `_fields_view_get` and `fields_view_get` are converted to a kwargs `**options`
  in `_get_view` and `get_view`.
  The rationale is that submenu was already no longer used (deprecated)
  and the mobile options is introduced.
  The mobile options is necessary to tell the server to send the mobile views
  for x2many fields (kanban instead of tree).
  Instead of adding a new argument each time we add a new option to
  `fields_view_get`, it seems wiser to have a kwargs `**options` to avoid
  to re-write all overrides each time a new option is introduced.
- `_fields_view_get` returned a dict containing the arch in text and some of the
  view information. Now, `get_view` returns a tuple with the view architecture
  as an `etree` node, and the view as a browse record. The rationale is that all
  overrides of `_fields_view_get` were about modifying the arch only
  (e.g. changing the address format/re-organizing the address related field
  nodes of the partner according to the company country).
  To do so, all these overrides were doing `etree.fromstring` to parse the arch
  which was sent in text to convert it to an `etree`,
  then operations were done on the `etree`,
  and then `etree.tostring` was called to convert back the arch to string.
  With this change of signature to send the arch as an `etree`,
  all these back and forth `etree.fromstring` -> `etree.tostring` are avoided,
  allowing some performance gain and less code in the end.
- A cleanup of the keys returned in the dict of `fields_view_get`
  has been performed in `get_view`:
  - `fields` is removed, as explained above,
  - `view_id` is renamed `id`,
  - `name` is removed, it was unused by the web client,
  - `type` is removed, it was unused by the web client,
  - `field_parent` is removed, it was unused by the web client,
  - `base_model` is removed, it was unused by the web client.
- `filters` is moved from the global dict returned by `load_views`
  (now `get_views`) to the dict returned by `fields_view_get` (now `get_view`)
  as it applies only to the `search` view type.
- Retro-compatible methods for the 3 methods
  `fields_view_get`, `_fields_view_get` and `load_views` are provided,
  with deprecation warnings in them.

# Future quests

- The web client could cache the model fields description
  (as it already caches the views),
  so it doesn't need to fetch them again if it asks for another view of a model
  for which he already has the fields description.
  If we do so, `get_views` could return only the list of models used by
  the views, without the fields description as of now,
  and the web client would then call `fields_get` independently only for
  the models for which it doesn't have yet the fields description.
  This would avoid the server to return the fields description
  and to call `fields_get`, which is costly, for each `get_views`,
  therefore gaining performances.
- Inject the views of the one2many and many2many fields all the way down,
  unlimited depth level, as explained above.
- Cache with `ormcache` the architecture of back-end views.
  This is already done for qweb views, it's not done for back-end views.
  Therefore the postprocessing of the views is performed for each `get_views`,
  which is costly, while the view architecture doesn't change for users
  belonging to the same groups, according to the groups implied by the view.

### Credits
This pull request is co-authored by
Aaron Bohy (aab) for the web client part and
Denis Ledoux (dle) for the server part.