Python's reference cycle collector

The standard implementation of CPython 3 has both reference counting and a generational garbage collector. The generational garbage collector is responsible for collecting reference cycles periodically. To review, here is an example of a reference cycle (note the use of .append).

x = [1, 2, 3]
y = [x]
x.append(y)
print(x)
# [1, 2, 3, [[...]]]
# [...] is a tell-tale sign of a ref cycle

Let’s see the generational garbage collector in action. Define a few helper functions for our investigation.

import sys, os, gc, psutil  # ignore E401
import numpy as np

process = psutil.Process(os.getpid())

def print_memory():
    memory_mb = int(np.round(process.memory_info().rss / 1e6))
    print('{}MB'.format(memory_mb))

def create_ref_cycle():
    a = [np.random.rand(2000, 2000)]
    b = [a]
    a.append(b)  # using .append() creates a ref cycle

def no_create_ref_cycle():
    a = [np.random.rand(2000, 2000)]
    b = [a]
    a = a + [b]  # redefining `a` doesn't create a ref cycle

Garbage collector is enabled by default. Let’s disable the garbage collector and create lots of reference cycles that become inaccessible to us. Measuring memory usage is a messy business that depends on lots of factors including the OS type. We will look at the resident set (rss) memory which may not be what your OS’s GUI reports (such as the Activity Monitor in MacOS). Your results will vary wildly every time you run these snippets. Our aim is to notice the effect of enabling or disabling the garbage collection instead of trying to get accurate memory usage. Also note that Python itself doesn’t always return all of the unused memory back to the OS.

# In a new python session with helper functions defined
gc.disable()
print_memory()
# 63MB
for _ in range(100):
    create_ref_cycle()  # consumes ~32MB per iteration
    print_memory()
# ...
# 3168MB
# 3200MB
# 3232MB
# 3264MB
gc.enable()
gc.collect()
print_memory()
# 2112MB

Let’s now run the function that does not create reference cycles. The memory usage for this function does not keep on increasing as a result of reference cycles (though there might be minor increases due to other, unrelated reasons).

# In a new python session with helper functions defined
gc.disable()
print_memory()
# 63MB
for _ in range(100):
    no_create_ref_cycle()  # consumes nothing per iteration
    print_memory()
# ...
# 95MB
# 95MB
# 95MB
# 95MB
gc.enable()
gc.collect()
print_memory()
# 95MB

The absolute memory numbers are not reliable – they vary by OS, RAM, other processes, and lots of other factors. Garbage collection runs periodically (not continuously) based on heuristics. You may see a temporary increase in memory usage until the garbage collection runs again. GC is also not perfect and does not prevent all memory leaks.

Takeaway

Don’t create reference cycles in your code. Know that the garbage collection in python is imperfect.