Python: Memory leak debugging

Question

I have a small multithreaded script running in django and over time its starts using more and more memory. Leaving it for a full day eats about 6GB of RAM and I start to swap.

Following http://www.lshift.net/blog/2008/11/14/tracing-python-memory-leaks I see this as the most common types (with only 800M of memory used):

(Pdb)  objgraph.show_most_common_types(limit=20)
dict                       43065
tuple                      28274
function                   7335
list                       6157
NavigableString            3479
instance                   2454
cell                       1256
weakref                    974
wrapper_descriptor         836
builtin_function_or_method 766
type                       742
getset_descriptor          562
module                     423
method_descriptor          373
classobj                   256
instancemethod             255
member_descriptor          218
property                   185
Comment                    183
__proxy__                  155

which doesn't show anything weird. What should I do now to help debug the memory problems?

Update: Trying some things people are recommending. I ran the program overnight, and when I work up, 50% * 8G == 4G of RAM used.

(Pdb) from pympler import muppy
(Pdb) muppy.print_summary()
                                     types |   # objects |   total size
========================================== | =========== | ============
                                   unicode |      210997 |     97.64 MB
                                      list |        1547 |     88.29 MB
                                      dict |       41630 |     13.21 MB
                                       set |          50 |      8.02 MB
                                       str |      109360 |      7.11 MB
                                     tuple |       27898 |      2.29 MB
                                      code |        6907 |      1.16 MB
                                      type |         760 |    653.12 KB
                                   weakref |        1014 |     87.14 KB
                                       int |        3552 |     83.25 KB
                    function (__wrapper__) |         702 |     82.27 KB
                        wrapper_descriptor |         998 |     77.97 KB
                                      cell |        1357 |     74.21 KB
  <class 'pympler.asizeof.asizeof._Claskey |        1113 |     69.56 KB
                       function (__init__) |         574 |     67.27 KB

That doesn't sum to 4G, nor really give me any big data structured to go fix. The unicode is from a set() of "done" nodes, and the list's look like just random weakrefs.

I didn't use guppy since it required a C extension and I didn't have root so it was going to be a pain to build.

None of the objectI was using have a __del__ method, and looking through the libraries, it doesn't look like django nor the python-mysqldb do either. Any other ideas?

Answer 1

Is DEBUG=False in settings.py?

If not Django will happily store all the SQL queries you make which adds up.

Answer 2

Do you use any extension? They are a wonderful place for memory leaks, and will not be tracked by python tools.

Answer 3

Have you tried gc.set_debug() ?

You need to ask yourself simple questions:

• Am I using objects with __del__ methods? Do I absolutely, unequivocally, need them?
• Can I get reference cycles in my code? Can't we break these circles before getting rid of the objects?

See, the main issue would be a cycle of objects containing __del__ methods:

import gc

class A(object):
    def __del__(self):
        print 'a deleted'
        if hasattr(self, 'b'):
            delattr(self, 'b')

class B(object):
    def __init__(self, a):
        self.a = a
    def __del__(self):
        print 'b deleted'
        del self.a


def createcycle():
    a = A()
    b = B(a)
    a.b = b
    return a, b

gc.set_debug(gc.DEBUG_LEAK)

a, b = createcycle()

# remove references
del a, b

# prints:
## gc: uncollectable <A 0x...>
## gc: uncollectable <B 0x...>
## gc: uncollectable <dict 0x...>
## gc: uncollectable <dict 0x...>
gc.collect()

# to solve this we break explicitely the cycles:
a, b = createcycle()
del a.b

del a, b

# objects are removed correctly:
## a deleted
## b deleted
gc.collect()

I would really encourage you to flag objects / concepts that are cycling in your application and focus on their lifetime: when you don't need them anymore, do we have anything referencing it?

Even for cycles without __del__ methods, we can have an issue:

import gc

# class without destructor
class A(object): pass

def createcycle():
    # a -> b -> c 
    # ^         |
    # ^<--<--<--|
    a = A()
    b = A()
    a.next = b
    c = A()
    b.next = c
    c.next = a
    return a, b, b

gc.set_debug(gc.DEBUG_LEAK)

a, b, c = createcycle()
# since we have no __del__ methods, gc is able to collect the cycle:

del a, b, c
# no panic message, everything is collectable:
##gc: collectable <A 0x...>
##gc: collectable <A 0x...>
##gc: collectable <dict 0x...>
##gc: collectable <A 0x...>
##gc: collectable <dict 0x...>
##gc: collectable <dict 0x...>
gc.collect()

a, b, c = createcycle()

# but as long as we keep an exterior ref to the cycle...:
seen = dict()
seen[a] = True

# delete the cycle
del a, b, c
# nothing is collected
gc.collect()

If you have to use "seen"-like dictionaries, or history, be careful that you keep only the actual data you need, and no external references to it.

I'm a bit disappointed now by set_debug, I wish it could be configured to output data somewhere else than to stderr, but hopefully that should change soon.

Answer 4

Try Guppy.

Basicly, you need more information or be able to extract some. Guppy even provides graphical representation of data.

Answer 5

I think you should use different tools. Apparently, the statistics you got is only about GC objects (i.e. objects which may participate in cycles); most notably, it lacks strings.

I recommend to use Pympler; this should provide you with more detailed statistics.

Answer 6

See this excellent blog post from Ned Batchelder on how they traced down real memory leak in HP's Tabblo. A classic and worth reading.