Skip to content

A runtime python code manipulation framework for profiling, debugging and bugfixing

License

Notifications You must be signed in to change notification settings

devopspp/pyliveupdate

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

64 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PyLiveUpdate

PyLiveUpdate is tool to help you modify your running python code without stopping it. It is useful for modifying long-running server programs in production with zero downtime. Some modification scenario includes: inject code to profile the runtime of certain lines of code; inject code to print out a variable for debugging; apply a patch without restarting.

Install

pip install pyliveupdate

How to use

Put these two lines of code in your server program and run it. Try it out with example/program1.py.

from pyliveupdatescripts import UpdateStub
UpdateStub().start()

Start a controller to modify it!

> pylu-controller -l 127.0.0.1:50050

Some predefined modification available in the controller

> FuncProfiler.profile(['__main__.**', 'module1.**']) # inject execution time profiling code into all functions in __main__ and module1
> LineProfiler.profile('__main__.bar', [11, 12]) # inject time profiling code into certain lines
> FuncDebugger.debug('__main__.bar') # inject code to print out function parameter and return value
> LineDebugger.debug('__main__.bar', [11, 12]) # inject code to print out variables in certain lines
> VarDebugger.debug('__main__.bar', 'b') # inject code to print out all values of a variable in different statements

These will print out corresponding profiling and debugging info in the program terminal and in /tmp/pyliveupdate. You can also define your own customized modifications.

Customized modification

There are in general two kinds of modification: instrument and redefine. You can define them as following and apply with patch('patch.py'). Try it out with example/patch.py.

Instrument code into existing functions

from pyliveupdate.update import Instrument, UpdateManager

def _line_after(a):
    print(a)
    
update = Instrument('__main__.bar', 
                    {('line_after', [12, 14]): _line_after})
UpdateManager.apply_update(update)

The code injects a print(a) in line 12 and 14 in function __main__.bar.

Redefine (patch) existing functions

from pyliveupdate.update import Redefine, UpdateManager

def new_bar(a):
    print('new_bar', a)
    
update = Redefine('__main__', None, {'__main__.bar':new_bar})
UpdateManager.apply_update(update)

The code redefines __main__.bar with new_bar.

Revert the modification

PyLiveUpdate also support to revert a modification on the fly:

> LU.ls() # list all modification
> LU.revert(1) # revert modifation with id 1

Extended tools

PyLiveUpdate also comes with some handy tools based on it:

Profiler

Dynamically choose functions or lines to profile and show the result with a flamegraph.

  1. Start your program with PyLiveUpdate enabled, like example/patch.py.
  2. Start a controller pylu-controller -l 127.0.0.1:50050
  3. Start profiling functions with FP.profile(['__main__.**', 'module1.**']) or lines with LineProfiler.profile('__main__.bar', [11, 12]).
  4. Process the logs with pylu-processlogs -i /tmp/pyliveupdate.log in another terminal It will generated a flamegraph and a summary:

Flamegraph

alt text

Function call summary

The following summary gives in process 4510 thread 5, views.results was called 10 times and each time takes 139 ms, views.results called manager.all for 20 times.

4510-Thread-5
function  hit  time/hit (ms)
views.results 10  138.562
  -manager.all 20  14.212
    -__init__.__hash__ 10  0.035
    -manager.get_queryset 20  0.922
      -query.__init__ 20  0.616
        -query.__init__ 20  0.071

Citation

Please kindly cite our work as follows if you want to refer to it in any writing:

@inproceedings{huang2021pylive,
  title={$\{$PYLIVE$\}$:$\{$On-the-Fly$\}$ Code Change for Python-based Online Services},
  author={Huang, Haochen and Xiang, Chengcheng and Zhong, Li and Zhou, Yuanyuan},
  booktitle={2021 USENIX Annual Technical Conference (USENIX ATC 21)},
  pages={349--363},
  year={2021}
}