Reduce memory footprint of server
The tricks used are:
- use a global LRU for `ormcache` (auto-balancing between registries);
- remove unnecessary data structures (binary class hierarchy of models);
- compute some data structures on demand (`_all_columns`);
- optimize field attributes (empty collections may be shared);
- optimize memory size of fields and columns by using slots.
On a database with modules sale, purchase and stock installed, the memory footprint of the registry went from 20.3Mb to 7.4Mb (as measured with heapy). In other words, the memory footprint was reduced to 1/3 !
In ormcache, those statistic numbers were counted per cached method for all
registries. We introduce separate statistic counters, and create a counter per
(database, model, method). The existing attributes that are no longer used
have been removed.
The registry size is now assumed to be around 10Mb, and ormcache size is
proportional to the maximum number of registries. Statistics about ormcache
usage is shown to the log when receiving signal SIGUSR1.
In `__getattr__`, remove the test on accessing `_attrs`. This situation should
never happen, since the attribute `_attrs` is set in method `__init__()`.
Introduce slots on all field classes for common attributes; slots take much
less memory than a `__dict__`. The other attributes are stored in a dictionary
`_attrs`; all fields with an empty value for `_attrs` (common case) share the
same empty dictionary. This saves quite some memory (around 4.5Mb per
registry), given the number of field instances created for a registry.
Another mechanism is used for the default values of attributes, since slots
cannot be assigned on classes.
The optimization consists in using tuples for attributes `inverse_fields`,
`computed_fields` and `_triggers`, and to let them share their value when it is
empty, which is common. This saves around 1.8Mb per registry.
This saves about 400Kb per registry by not allocating those lists, which are in
most cases empty. The removal of the attribute will also simplify a bit the
management of free attributes.
The computed value of parameter digits is no longer stored into fields and
columns; instead the value is recomputed everytime it is needed. Note that
performance is not an issue, since the method `get_precision` of model
'decimal.precision' is cached by the orm. This simplifies the management of
digits on fields and saves about 300Kb per registry.
Sometimes, the expected mro of the model is not the same as the one built with
a binary class hierarchy. So we reorder the base classes in order to match the
equivalent binary class hierarchy. This also fixes the cases where duplicates
appear in base classes.
Instead of composing classes in a binary tree hierarchy, we make one class that
inherits from all its base classes. This avoids keeping intermediate data
structures for columns, inherits, constraints, etc. This saves about 600Kb per
registry.
The mappings model._all_columns takes quite some memory (around 2MB per
registry) because of the numerous dictionaries (one per model) and inefficient
memory storage of column_info. Since it is deprecated and almost no longer
used, it can be computed on demand.
The ormcache is now shared among registries. The cached methods use keys like
(DBNAME, MODELNAME, METHOD, args...). This allows registries with high load to
use more cache than other registries.
QWeb templates can be translated, javascript code
included. Nothing prevents you to translate
the javascript reserved word "undefined", for instance.
Preventing the livechat JS snippet to be translated
prevent translations mistakes, like translate
```undefined``` by ```Niet gedefinieerd```
opw-630335
The datetime widget use the "keypress" event to detect if it has been
changed. Unfortunately, some keys, such as backspace, do not fire a
keypress event. This commit listen instead on the 'change' event, which
should fire when the input has changed.
The same holds for the inverse history_ids relationship. When an object is copied,
the many2many fields their links are copied.
When we copy a done move as is done in the return wizard to create the reverse, it copied also
the relationship with the quants. This is problematic as this field indicates the quants that
were transferred by the move and the new move will think it will have returned all the quants
even before it is done.
Some Belgian banks (e.g. CPH) seem to generate CODA
import files that do not comply with the CODA 2.3
(or 2.4 or 2.5) spec, specificially in the header
we use to detect CODA files.
opportunities button tip. Using directly the first child
is risky because modules can add statbutton before the
opportunities button (published for example). Using a
class seems a safer way to select the target.
Currently, the view_form search_count doesn't propagate the context, so when
searching on a translated field, the count can be different than the one we
would expect and get with a search in a current language different than en_US.
opw-628792 opw-630212 closes#5825
The displayed date was incorrect: instead of the day number in the month (1st-31th), we displayed the day order in the week (0th-6th).
Yeah, this one was a tricky one...
Richard Mathot