Django update settings in Celery's async task

Question

I am using celery to load Neural Network models and would like to store the loaded model in settings for fast prediction.

so in django.conf.settings I have:

MODELS = {}

and in celery task, I have the following snippet:

@app.task
def load_nn_models(model_name):
     from django.conf import settings
     ...    
     settings.MODELS[model_name] = {'model': net, 'graph': sess}

However, I noticed that the tasks are running in another thread that launches different Django Environment and any changes in the settings will not be reflected back to the main thread.

Is there a workaround for this?

EDIT

The parameters I am storing in settings are:

net: keras Model
sess: Tensorflow Session

@BearBrown I don't think it will work because when django starts, it reads the settings into dictionary and keep the changes on that dictionary not on a file. so if main django changed something the async task will not be able to get it. — Coderji, Sep 06 '18 at 11:42

score 2 · Answer 1 · answered Sep 06 '18 at 12:21

2

Django settings are not the right place for this, obviously. First because the settings object is not a shared resource (there's one instance per process), then because the doc explicitely mentions that this object is to be considered as immutable.

If your point is to have a celery task computing those objects so that other tasks and / or the front can use them, you will have to find a way to serialize them and store the serialized version in a shared resource (database, cache, etc).

answered Sep 06 '18 at 12:21

bruno desthuilliers

75,974
6
88
118

I agree with you. But the only way I can think of is to store nn models as binary streams then because putting them in a shared resource such as db or file will take too much time to reload the model and do prediction for every API request. – Coderji Sep 06 '18 at 12:48
Well this is a completely different question anyway, and you really didn't provide enough context for someone to even give a hint. I suggest you post a new question explaining the real problem instead of "what you think is the solution" (XY problem as usual), making sure to provide all the relevant informations (if those objects are per-user or global to the site, how often they change, what triggers a change, how "resource-heavy" they are - memory size etc - etc etc etc). Also make sure you've explored those libs (keras an tf) in depth, sometimes there are half-hidden features. – bruno desthuilliers Sep 06 '18 at 13:32

Brown Bear · Answer 2 · 2018-09-06T12:10:32.947

1

you can try to use the configparser import configparser

def dict_from_file():
    config = configparser.ConfigParser()
    config.read("config.ini")
    models = config['models']
    for x in models.values():
        print(x)

set file config.ini:

[models]
var_a: home
var_b: car
var_c: Next

call dict_from_file the output is:

home
car
Next

update the file config.ini:

[models]
var_a: home
var_c: New

call dict_from_file the output is:

home
New

ypu can read more for the supported-datatypes

edited Sep 06 '18 at 12:10

answered Sep 06 '18 at 12:00

Brown Bear

19,655
10
58
76

I see now, can `home`, `car` be any kind of object? – Coderji Sep 06 '18 at 12:06
@Coderji what do you mean object? django model or python dict? add link of suported types. – Brown Bear Sep 06 '18 at 12:09
for the extra objects i can't try and test but i think you can try youself with `yaml` by the same logic load info from file – Brown Bear Sep 06 '18 at 12:21

Django update settings in Celery's async task

2 Answers2