Django Celery task on Heroku causes high memory usage


I have a celery task on Heroku that connects to an external API and retrieves some data, stores in the database and repeats several hundred times. Very quickly (after ~10 loops) Heroku starts warning about high memory usage. Any ideas?

def retrieve_details():
    for p in PObj.objects.filter(some_condition=True):

def fetch(self):
    v_data = self.service.getV(**dict(
    response = self.map_response(v_data)

    for key in ["some_key","some_other_key",]:
        setattr(self.v, key, response.get(key))

Heroky logs

2017-01-01 10:26:25.634
132 <45>1 2017-01-01T10:26:25.457411+00:00 heroku run.5891 - - Error R14 (Memory quota exceeded)

Go to the log:

You are receiving this email because your Logentries alarm "Memory quota exceeded"
has been triggered.

In context:
2017-01-01 10:26:25.568 131 <45>1 2017-01-01T10:26:25.457354+00:00 heroku run.5891 - - Process running mem=595M(116.2%)
2017-01-01 10:26:25.634 132 <45>1 2017-01-01T10:26:25.457411+00:00 heroku run.5891 - - Error R14 (Memory quota exceeded)

Show source
| python   | django   | heroku   | django-celery   2017-01-01 11:01 2 Answers

Answers ( 2 )

  1. 2017-01-02 05:01

    You're basically loading a bunch of data into a Python dictionary in memory. This will cause a lot of memory overhead, especially if you are grabbing a lot of objects from the local database.

    Do you really need to store all of these objects in a dictionary?

    What most people do for things like this is:

    • Retrieve one object at a time from the database.
    • Process that item (perform whatever logic you need).
    • Repeat.

    This way, you only end up storing a single object in memory at any given time, thereby greatly reducing your memory footprint.

    If I were you, I'd look for ways to move my logic into the database query, or simply process each item individually.

  2. 2017-01-07 19:01

    To expand on the veritable rdegges thoughts, here are two strategies I have used in the past when working with celery/python to help reduce the memory footprint: (1) kick off subtasks that each process exactly one object and/or (2) use generators.

    1. kick off subtasks that each process exactly one object:

      def retrieve_details():
          qs = PObj.objects.filter(some_condition=True)
          for p in qs.values_list('id', flat=True):
      def do_fetch(n_id):
          p = PObj.objects.get(id=n_id)

      Now you can tune celery so that it kills of processes after processing N number of PObj (tasks) to keep memory footprint low using --max-tasks-per-child.

    2. Using generators: you can also try this using generators so that you can (theoretically) throw away the PObj after you call fetch

      def ps_of_interest(chunk=10):
          n = chunk
          start = 0
          while n == chunk:
              some_ps = list(PObj.objects.filter(some_condition=True)[start:start + n])
              n = len(some_ps)
              start += chunk
              for p in some_ps:
                  yield p
      def retrieve_details():
          for p in ps_of_interest():

    For my money, I’d go with option #1.

◀ Go back