Awesome

mem_top

Shows top suspects for memory leaks in your Python program.

Usage

pip install mem_top
from mem_top import mem_top

# From time to time:
logging.debug(mem_top())
# print(mem_top())

# Notice which counters keep increasing over time - they are the suspects.

Counters

mem_top iterates all objects found in memory and calculates:

• refs - number of direct references from this object to other objects, like keys and values of dict
  • E.g. a dict {("some", "complex", "key"): "value"} will have refs: 2 - 1 ref for key, 1 ref for value
  • Its key ("some", "complex", "key") will have refs: 3 - 1 ref per item
• bytes - size of this object in bytes
• types - number of objects of this type still kept in memory after garbage collection

Real life example

refs:
144997  <type 'collections.defaultdict'> defaultdict(<type 'collections.deque'>, {<GearmanJobRequest task='...', unique='.
144996  <type 'dict'> {'.:..............:.......': <GearmanJobRequest task='..................', unique='.................
18948   <type 'dict'> {...
1578    <type 'dict'> {...
968     <type 'dict'> {...
968     <type 'dict'> {...
968     <type 'dict'> {...
767     <type 'list'> [...
726     <type 'dict'> {...
608     <type 'dict'> {...

types:
292499  <type 'dict'>
217912  <type 'collections.deque'>
72702   <class 'gearman.job.GearmanJob'>
72702   <class 'gearman.job.GearmanJobRequest'>
12340   <type '...
3103    <type '...
1112    <type '...
855     <type '...
767     <type '...
532     <type '...

• Noticed a leak of 6GB RAM and counting.
• Added "mem_top" and let it run for a while.
• When got the result above it became absolutely clear who is leaking here:
  the Python client of Gearman kept increasing its counters over time.
• Found its known bug - https://github.com/Yelp/python-gearman/issues/10
  leaking defaultdict of deques, and a dict of GearmanJobRequest-s,
  just as the "mem_top" showed.
• Replaced "python-gearman" - long story: stale 2.0.2 at PyPI, broken 2.0.X at github, etc.
• "mem_top" confirmed the leak is now completely closed.

Updates

• Pass e.g. verbose_types=[dict, list] to store their values, sorted by repr length descending, in verbose_file_name or returned from mem_top().
• Added "bytes" top.

Config defaults

mem_top(
    limit=10,                           # limit of top lines per section
    width=100,                          # width of each line in chars
    sep='\n',                           # char to separate lines with
    refs_format='{num}\t{type} {obj}',  # format of line in "refs" section
    bytes_format='{num}\t {obj}',       # format of line in "bytes" section
    types_format='{num}\t {obj}',       # format of line in "types" section
    verbose_types=None,                 # list of types to sort values by `repr` length
    verbose_file_name='/tmp/mem_top',   # name of file to store verbose values in
)

See also

• https://docs.python.org/2/library/gc.html#gc.garbage
• https://pypi.python.org/pypi/objgraph

About

mem_top version 0.2.1
Copyright (c) 2014-2022 Denis Ryzhkov denisr@denisr.com
MIT License