skip to navigation
skip to content

Python Wiki

Python Insider Blog

Python 2 or 3?

Help Fund Python

[Python resources in languages other than English]

Non-English Resources

Add an event to this calendar.

Times are shown in UTC/GMT.

Add an event to this calendar.

PEP:454
Title:Add a new tracemalloc module to trace Python memory allocations
Version:d617c7ba4e14
Last-Modified:2016-05-03 13:52:22 +0300 (Tue, 03 May 2016)
Author:Victor Stinner <victor.stinner at gmail.com>
BDFL-Delegate:Charles-François Natali <cf.natali@gmail.com>
Status:Final
Type:Standards Track
Content-Type:text/x-rst
Created:3-September-2013
Python-Version:3.4
Resolution:https://mail.python.org/pipermail/python-dev/2013-November/130491.html

Abstract

This PEP proposes to add a new tracemalloc module to trace memory blocks allocated by Python.

Rationale

Classic generic tools like Valgrind can get the C traceback where a memory block was allocated. Using such tools to analyze Python memory allocations does not help because most memory blocks are allocated in the same C function, in PyMem_Malloc() for example. Moreover, Python has an allocator for small objects called "pymalloc" which keeps free blocks for efficiency. This is not well handled by these tools.

There are debug tools dedicated to the Python language like Heapy Pympler and Meliae which lists all alive objects using the garbage collector module (functions like gc.get_objects(), gc.get_referrers() and gc.get_referents()), compute their size (ex: using sys.getsizeof()) and group objects by type. These tools provide a better estimation of the memory usage of an application. They are useful when most memory leaks are instances of the same type and this type is only instantiated in a few functions. Problems arise when the object type is very common like str or tuple, and it is hard to identify where these objects are instantiated.

Finding reference cycles is also a difficult problem. There are different tools to draw a diagram of all references. These tools cannot be used on large applications with thousands of objects because the diagram is too huge to be analyzed manually.

Proposal

Using the customized allocation API from PEP 445, it becomes easy to set up a hook on Python memory allocators. A hook can inspect Python internals to retrieve Python tracebacks. The idea of getting the current traceback comes from the faulthandler module. The faulthandler dumps the traceback of all Python threads on a crash, here is the idea is to get the traceback of the current Python thread when a memory block is allocated by Python.

This PEP proposes to add a new tracemalloc module, a debug tool to trace memory blocks allocated by Python. The module provides the following information:

  • Traceback where an object was allocated
  • Statistics on allocated memory blocks per filename and per line number: total size, number and average size of allocated memory blocks
  • Computed differences between two snapshots to detect memory leaks

The API of the tracemalloc module is similar to the API of the faulthandler module: enable() / start(), disable() / stop() and is_enabled() / is_tracing() functions, an environment variable (PYTHONFAULTHANDLER and PYTHONTRACEMALLOC), and a -X command line option (-X faulthandler and -X tracemalloc). See the documentation of the faulthandler module.

The idea of tracing memory allocations is not new. It was first implemented in the PySizer project in 2005. PySizer was implemented differently: the traceback was stored in frame objects and some Python types were linked the trace with the name of object type. PySizer patch on CPython adds an overhead on performances and memory footprint, even if the PySizer was not used. tracemalloc attachs a traceback to the underlying layer, to memory blocks, and has no overhead when the module is not tracing memory allocations.

The tracemalloc module has been written for CPython. Other implementations of Python may not be able to provide it.

API

To trace most memory blocks allocated by Python, the module should be started as early as possible by setting the PYTHONTRACEMALLOC environment variable to 1, or by using -X tracemalloc command line option. The tracemalloc.start() function can be called at runtime to start tracing Python memory allocations.

By default, a trace of an allocated memory block only stores the most recent frame (1 frame). To store 25 frames at startup: set the PYTHONTRACEMALLOC environment variable to 25, or use the -X tracemalloc=25 command line option. The set_traceback_limit() function can be used at runtime to set the limit.

Functions

clear_traces() function:

Clear traces of memory blocks allocated by Python.

See also stop().

get_object_traceback(obj) function:

Get the traceback where the Python object obj was allocated. Return a Traceback instance, or None if the tracemalloc module is not tracing memory allocations or did not trace the allocation of the object.

See also gc.get_referrers() and sys.getsizeof() functions.

get_traceback_limit() function:

Get the maximum number of frames stored in the traceback of a trace.

The tracemalloc module must be tracing memory allocations to get the limit, otherwise an exception is raised.

The limit is set by the start() function.

get_traced_memory() function:

Get the current size and maximum size of memory blocks traced by the tracemalloc module as a tuple: (size: int, max_size: int).

get_tracemalloc_memory() function:

Get the memory usage in bytes of the tracemalloc module used to store traces of memory blocks. Return an int.

is_tracing() function:

True if the tracemalloc module is tracing Python memory allocations, False otherwise.

See also start() and stop() functions.

start(nframe: int=1) function:

Start tracing Python memory allocations: install hooks on Python memory allocators. Collected tracebacks of traces will be limited to nframe frames. By default, a trace of a memory block only stores the most recent frame: the limit is 1. nframe must be greater or equal to 1.

Storing more than 1 frame is only useful to compute statistics grouped by 'traceback' or to compute cumulative statistics: see the Snapshot.compare_to() and Snapshot.statistics() methods.

Storing more frames increases the memory and CPU overhead of the tracemalloc module. Use the get_tracemalloc_memory() function to measure how much memory is used by the tracemalloc module.

The PYTHONTRACEMALLOC environment variable (PYTHONTRACEMALLOC=NFRAME) and the -X tracemalloc=NFRAME command line option can be used to start tracing at startup.

See also stop(), is_tracing() and get_traceback_limit() functions.

stop() function:

Stop tracing Python memory allocations: uninstall hooks on Python memory allocators. Clear also traces of memory blocks allocated by Python

Call take_snapshot() function to take a snapshot of traces before clearing them.

See also start() and is_tracing() functions.

take_snapshot() function:

Take a snapshot of traces of memory blocks allocated by Python. Return a new Snapshot instance.

The snapshot does not include memory blocks allocated before the tracemalloc module started to trace memory allocations.

Tracebacks of traces are limited to get_traceback_limit() frames. Use the nframe parameter of the start() function to store more frames.

The tracemalloc module must be tracing memory allocations to take a snapshot, see the start() function.

See also the get_object_traceback() function.

Filter

Filter(inclusive: bool, filename_pattern: str, lineno: int=None, all_frames: bool=False) class:

Filter on traces of memory blocks.

See the fnmatch.fnmatch() function for the syntax of filename_pattern. The '.pyc' and '.pyo' file extensions are replaced with '.py'.

Examples:

  • Filter(True, subprocess.__file__) only includes traces of the subprocess module
  • Filter(False, tracemalloc.__file__) excludes traces of the tracemalloc module
  • Filter(False, "<unknown>") excludes empty tracebacks

inclusive attribute:

If inclusive is True (include), only trace memory blocks allocated in a file with a name matching filename_pattern at line number lineno.

If inclusive is False (exclude), ignore memory blocks allocated in a file with a name matching filename_pattern at line number lineno.

lineno attribute:

Line number (int) of the filter. If lineno is None, the filter matches any line number.

filename_pattern attribute:

Filename pattern of the filter (str).

all_frames attribute:

If all_frames is True, all frames of the traceback are checked. If all_frames is False, only the most recent frame is checked.

This attribute is ignored if the traceback limit is less than 2. See the get_traceback_limit() function and Snapshot.traceback_limit attribute.

Frame

Frame class:

Frame of a traceback.

The Traceback class is a sequence of Frame instances.

filename attribute:

Filename (str).

lineno attribute:

Line number (int).

Snapshot

Snapshot class:

Snapshot of traces of memory blocks allocated by Python.

The take_snapshot() function creates a snapshot instance.

compare_to(old_snapshot: Snapshot, group_by: str, cumulative: bool=False) method:

Compute the differences with an old snapshot. Get statistics as a sorted list of StatisticDiff instances grouped by group_by.

See the statistics() method for group_by and cumulative parameters.

The result is sorted from the biggest to the smallest by: absolute value of StatisticDiff.size_diff, StatisticDiff.size, absolute value of StatisticDiff.count_diff, Statistic.count and then by StatisticDiff.traceback.

dump(filename) method:

Write the snapshot into a file.

Use load() to reload the snapshot.

filter_traces(filters) method:

Create a new Snapshot instance with a filtered traces sequence, filters is a list of Filter instances. If filters is an empty list, return a new Snapshot instance with a copy of the traces.

All inclusive filters are applied at once, a trace is ignored if no inclusive filters match it. A trace is ignored if at least one exclusive filter matchs it.

load(filename) classmethod:

Load a snapshot from a file.

See also dump().

statistics(group_by: str, cumulative: bool=False) method:

Get statistics as a sorted list of Statistic instances grouped by group_by:

group_by description
'filename' filename
'lineno' filename and line number
'traceback' traceback

If cumulative is True, cumulate size and count of memory blocks of all frames of the traceback of a trace, not only the most recent frame. The cumulative mode can only be used with group_by equals to 'filename' and 'lineno' and traceback_limit greater than 1.

The result is sorted from the biggest to the smallest by: Statistic.size, Statistic.count and then by Statistic.traceback.

traceback_limit attribute:

Maximum number of frames stored in the traceback of traces: result of the get_traceback_limit() when the snapshot was taken.

traces attribute:

Traces of all memory blocks allocated by Python: sequence of Trace instances.

The sequence has an undefined order. Use the Snapshot.statistics() method to get a sorted list of statistics.

Statistic

Statistic class:

Statistic on memory allocations.

Snapshot.statistics() returns a list of Statistic instances.

See also the StatisticDiff class.

count attribute:

Number of memory blocks (int).

size attribute:

Total size of memory blocks in bytes (int).

traceback attribute:

Traceback where the memory block was allocated, Traceback instance.

StatisticDiff

StatisticDiff class:

Statistic difference on memory allocations between an old and a new Snapshot instance.

Snapshot.compare_to() returns a list of StatisticDiff instances. See also the Statistic class.

count attribute:

Number of memory blocks in the new snapshot (int): 0 if the memory blocks have been released in the new snapshot.

count_diff attribute:

Difference of number of memory blocks between the old and the new snapshots (int): 0 if the memory blocks have been allocated in the new snapshot.

size attribute:

Total size of memory blocks in bytes in the new snapshot (int): 0 if the memory blocks have been released in the new snapshot.

size_diff attribute:

Difference of total size of memory blocks in bytes between the old and the new snapshots (int): 0 if the memory blocks have been allocated in the new snapshot.

traceback attribute:

Traceback where the memory blocks were allocated, Traceback instance.

Trace

Trace class:

Trace of a memory block.

The Snapshot.traces attribute is a sequence of Trace instances.

size attribute:

Size of the memory block in bytes (int).

traceback attribute:

Traceback where the memory block was allocated, Traceback instance.

Traceback

Traceback class:

Sequence of Frame instances sorted from the most recent frame to the oldest frame.

A traceback contains at least 1 frame. If the tracemalloc module failed to get a frame, the filename "<unknown>" at line number 0 is used.

When a snapshot is taken, tracebacks of traces are limited to get_traceback_limit() frames. See the take_snapshot() function.

The Trace.traceback attribute is an instance of Traceback instance.

Rejected Alternatives

Log calls to the memory allocator

A different approach is to log calls to malloc(), realloc() and free() functions. Calls can be logged into a file or send to another computer through the network. Example of a log entry: name of the function, size of the memory block, address of the memory block, Python traceback where the allocation occurred, timestamp.

Logs cannot be used directly, getting the current status of the memory requires to parse previous logs. For example, it is not possible to get directly the traceback of a Python object, like get_object_traceback(obj) does with traces.

Python uses objects with a very short lifetime and so makes an extensive use of memory allocators. It has an allocator optimized for small objects (less than 512 bytes) with a short lifetime. For example, the Python test suites calls malloc(), realloc() or free() 270,000 times per second in average. If the size of log entry is 32 bytes, logging produces 8.2 MB per second or 29.0 GB per hour.

The alternative was rejected because it is less efficient and has less features. Parsing logs in a different process or a different computer is slower than maintaining traces on allocated memory blocks in the same process.

Prior Work

  • Python Memory Validator (2005-2013): commercial Python memory validator developed by Software Verification. It uses the Python Reflection API.
  • PySizer: Google Summer of Code 2005 project by Nick Smallbone.
  • Heapy (2006-2013): part of the Guppy-PE project written by Sverker Nilsson.
  • Draft PEP: Support Tracking Low-Level Memory Usage in CPython (Brett Canon, 2006)
  • Muppy: project developed in 2008 by Robert Schuppenies.
  • asizeof: a pure Python module to estimate the size of objects by Jean Brouwers (2008).
  • Heapmonitor: It provides facilities to size individual objects and can track all objects of certain classes. It was developed in 2008 by Ludwig Haehne.
  • Pympler (2008-2011): project based on asizeof, muppy and HeapMonitor
  • objgraph (2008-2012)
  • Dozer: WSGI Middleware version of the CherryPy memory leak debugger, written by Marius Gedminas (2008-2013)
  • Meliae: Python Memory Usage Analyzer developed by John A Meinel since 2009
  • gdb-heap: gdb script written in Python by Dave Malcom (2010-2011) to analyze the usage of the heap memory
  • memory_profiler: written by Fabian Pedregosa (2011-2013)
  • caulk: written by Ben Timby in 2012

See also Pympler Related Work.