memory_utils

Yeah Memory Issues!!

Memory Issues happen to the best of us. memory_utils will give you simple tools to quickly isolate the cuplrit, and ideally, warn you before you run into issues.

From my experience, there is no silver-bullet in dealing with memory issues. You just have to roll up your sleeve and get dirty with print statements. In our team's recent fight with a memory issue, we created memory_utils and we wanted to share.

memory_utils deals primarily with RSS memory (Resident Set Size). The most important memory concept to understand when dealing with memory constrained systems: RSS, the resident set size, is the portion of a process's memory that is held in RAM. The rest of the memory exists in the swap of the file system.

Install

pip install memory_utils

Usage

print_memory

The workhorse of this package is print_memory It simply prints out 3 columns of data: the current memory, the delta since the previous statement and an message that you pass it. If there is additional memory used -- the line will be printed RED and if there is a decrease, the line will be printed GREEN.

It is a very simple approach, but it really helped us find out where the issue was, at glance. The output could look like this:

RSS Delta Message
14,393,344 14,393,344 BEFORE BLOAT
14,397,440 4,096 DURING BLOAT (1)
14,413,824 16,384 DURING BLOAT (102)
14,417,920 4,096 DURING BLOAT (211)
14,438,400 20,480 DURING BLOAT (1002)
14,442,496 4,096 DURING BLOAT (2034)
14,462,976 20,480 DURING BLOAT (2056)

memory_watcher and check_memory

We have worker processes that run in containers. I like to fail hard and early. So we have two helper functions that help us with that

check_memory

Will check the current rss memory against the memory_utils set memory limit. And if it crosses that limit it will raise a MemoryTooBigException

pipinstallmemory_utilsimportmemory_utilsmemory_utils.set_memory_limit(200*memory_utils.MEGABYTES)
# .... else wherememory_utils.check_memory()

memory_watcher

Often you will want to do your check_memory at a _safe_ place. Also memory leaks often happen within a loop. We created memory_watcher with those concepts in mind

foraccountinmemory_watcher(Account.objects):
account.do_something_memory_intensive()
account.save()

This will call check_memory before each iteration

Configuration

set_verbose

By default print_memory will only print statements that move the memory and memory_watcher will not print its memory usage. If you want additional verbosity set this to true

importmemory_utilsmemory_utils.set_verbose(True)

set_memory_limit

By default, the memory limit at 200 MB.

Use this method to change the default.

This setting is used in print_memory and memory_watcher

Note: you can also override this limit at the function level as well

importmemory_utilsmemory_utils.set_memory_limit(500*memory_utils.MEGABYTES)

set_out

By default, we will print to standard out. Feel free to override here like so

importmemory_utilsfromStringIOimportStringIOout=StringIO()
memory_utils.set_out(out)

Questions / Issues

Feel free to ping me on twitter: @tushman or add issues or PRs at https://github.com/jtushman/memory_utils