RevDB, the Reverse Debugger


A "reverse debugger" is a debugger where you can go forward and backward in time. RevDB is a reverse debugger for Python based on PyPy. It can be used to track down hard-to-reproduce bugs in your Python programs (whether you usually run them on PyPy or on CPython).

It is similar to undodb-gdb and rr, which are reverse debuggers for C code. RevDB does not allow you to step in or inspect things at the level of C: it works purely on Python.

For more information about what reverse debugging is, have a look at https://en.wikipedia.org/wiki/Debugger#Reverse_debugging and http://programmers.stackexchange.com/questions/181527/why-is-reverse-debugging-rarely-used.

RevDB is tested on Linux and to some extent on OS/X. It is known not to work on Windows so far. It should support both Python 2.7 and 3.x as available in PyPy.

This is the original blog post, which describes the basics: https://morepypy.blogspot.ch/2016/07/reverse-debugging-for-python.html (note that many of the limitations described in that blog post have been removed now: threads and cpyext are implemented; various crashes have been fixed; next-style commands behave more reasonably now; import in ! commands is special-cased).


You need to download and build a special version of PyPy. Sorry, there are not prebuilt binaries at this point in time. This is mainly because distributing Linux binaries is a mess. Note that the building process takes 20 to 30 minutes (which is long, but less than a normal PyPy).

  • Get PyPy's source code: see http://pypy.org/download.html#building-from-source (point 1). Starting from PyPy v7.0.0 you don't need to use a special reverse-debugger branch.

  • Make sure you have the dependencies installed: http://pypy.readthedocs.org/en/latest/build.html#install-build-time-dependencies (note that for RevDB you need the Boehm garbage collector libgc, even if you don't plan to run PyPy's tests).

  • Build the revdb version of PyPy:

    cd pypy/goal
    ../../rpython/bin/rpython -O2 --revdb

    Known issue: on Linux kernel 4.12, address space layout randomization (ASLR) cannot be disabled any more. This is a blocker for RevDB. (It works if you build your revdb version with the --no-shared option, but this means you cannot use cpyext, the C extension module compatibility layer.) Kernel 4.19 works again. The situation for intermediate kernels is unknown.

  • Finally, you need to install the regular, almost-pure Python package https://bitbucket.org/pypy/revdb (which is where the present README file originally lives). It has got a small CFFI module, so you should run either python setup.py install (usually in a virtualenv) or directly python setup.py build_ext --inplace. Use a regular CPython 2.7.x here, or PyPy2 but not the RevDB version of PyPy. (You need to use Python 2.x for this even if you did a recording of a Python 3.x program above.)


  • Here is the executable you use instead of pypy or python:


    It works like a (slow but) regular Python interpreter, so you can make virtualenvs with it, or do any necessary preparation. You can even install CPython C extension modules, which will work with PyPy's cpyext support---which is, in itself, only a 99% solution: a few CPython C extension modules may not fully work in PyPy. But if they work in a regular PyPy they should work with the RevDB PyPy too.

  • Recording: When you are ready to run the program that you want to debug, use the REVDB environment variable to ask the above pypy-c to write a log file:

    REVDB=log.rdb  /.../pypy/goal/pypy-c  yourprogr.py  arguments...

    You can repeat the step above until you succeed in logging an execution that exhibits the bug that you are tracking. Once you do, you get a log.rdb that we will use next. The same log.rdb can be used any number of times for replaying. In case of doubt, if it was hard to obtain, make a safe copy.

  • Replaying: start the debugger's user interface:

    /path/to/revdb/revdb.py  log.rdb

    If you want to enable syntax coloring, add -c dark or -c light depending on whether you use a dark- or light-background terminal (you need to install pygments, then).

    Do not run this in the virtualenv you created in the previous step! This must run with a regular Python (CPython 2.7.x, or non-RevDB PyPy2).

    Replaying works by having revdb.py find the pypy-c of RevDB and internally executing it in a special mode. It looks at the path recorded in the log file (but see also the -x argument). It must find the very same version of pypy-c. With that restriction, you could in theory move that log.rdb file on another machine and debug there, if the pypy-c executable and associated libpypy-c.so work when copied unchanged on that machine too.

Note that the log file typically grows at a rate of 1-2 MB per second. Assuming size is not a problem, the limiting factor are:

  1. Replaying time. If your recorded execution took more than a few minutes, replaying will be painfully slow. It sometimes needs to go over the whole log several times in a single session. If the bug occurs randomly but rarely, you should run recording for a few minutes, then kill the process and try again, repeatedly until you get the crash.
  2. RAM usage for replaying. The RAM requirements are 10 or 15 times larger for replaying than for recording. If that is too much, you can try with a lower value for MAX_SUBPROCESSES in _revdb/process.py, but it will always be several times larger.

Debugger User Interface

The debugger user interface is a mix between gdb and pdb. Type "help" to get a summary of all commands.

(Write more here...)

You can get a feel for the commands by following the blog post https://morepypy.blogspot.ch/2016/07/reverse-debugging-for-python.html.

Below we give some description for the least obvious but most useful commands.


This is the prompt, which displays the current timestamp. The timestamp fully identifies the position in the log. Use go to jump directly to some timestamp number. Use step/bstep to do single-timestamp steps. Other commands step by more, like next and finish and their b variants.


This is usually the first command you give, to go to the last timestamp before stepping back. A breakpoint-like "stoppoint" is set automatically and is always present: it activates at the time when execution just finished running the main module. There are more recorded timestamps afterwards, particularly if PyPy is then going to print a traceback, but you are generally not interested in that. So after you start revdb.py you typically say continue, hit the stoppoint, and then say bstep a few times to reach the last interesting point (e.g. where the exception was raised, assuming there was one).

Note another trick, useful if running tests: it's hard to go to the correct place if the testing framework does a lot of extra things after the failure occurs. Then you can put os._exit(1) in your test instead of, say, the failing assert; and then when replaying, continue will go to that place.


The print command can run any Python code, including (single-line) statements. It only prints the result if it was an expression and that expression returns a result different from None. In other words, it works like typing at Python's interactive mode does; it does not work like Python's own print statement. It is sometimes clearer to use !, which is another abbreviation for print or p.

$5 =

Whenever a dynamic (i.e. non-prebuilt) object is printed, it is printed with a numeric prefix, e.g. $5 =. Afterwards, you can use the expression $5 in all Python expressions; it stands for the same object. The parser recognizes it as a standard subexpression, so you can say $5.foo or len($5) etc. It continues to work after you move at a different time in the past or the future. If you move before the time of creation for this object, using $5 will raise an exception. Note that the existence of $5 keeps the object alive forever (it can be recalled even if you go far in the future), but this doesn't change the recorded program's own results: the __del__ method is called, and weakrefs to $5 go away, as per the recording.


break puts a breakpoint, either by line number or by function name. If you say break foo or break foo() with empty parentheses, the breakpoint activates whenever a function with the name foo is called. To set a breakpoint by line number, use either break NUM or break FILE:NUM. The FILE defaults to the co_filename of the current code object. If given explicitly, FILE matches any code object with a co_filename of the form /any/path/FILE. For example, if you set a breakpoint at foo.py:42 it will break at the line 42 in any file called /any/path/foo.py. (Breakpoints cannot be conditional for now.)

nthread, bthread

Multithreaded programs are handled correctly. As usual with the GIL, in the recording session only one thread can run Python bytecodes at a time; so during replaying (i.e. now) you see bytecodes executed sequentially. revdb.py displays a marker line whenever the next place it displays is actually from a different thread than the last. Typically, thread switches occur rarely. You can use the nthread and bthread commands to go forward or backward until a thread switch occurs (either going to any different thread, or going precisely to the thread with the given number).


watch puts a watchpoint. This command is essential to RevDB's debugging approach! Watchpoints are expressions that are evaluated outside any context, so they must not depend on any local or global variable. They can depend on builtins, and they can use $NUM to reference any previously-printed object. Usually we watch $2.foo to find where the attribute foo on this precise object $2 changed; or len($3) to find where the length of the list $3 changed. Similarly, you can find out who changes the value of the global mod.GLOB: first do print mod to get $4 = <module...> and then set a watchpoint on $4.GLOB. It may occasionally be useful to set a watchpoint on just $5: it means that you're watching for changes in the repr of this exact object.

If you are a bit creative you can call a Python function from your program: first print the function itself, and then set a watchpoint on, say, $6() > 100. However, watchpoint expressions must be fully side-effect-free, otherwise replaying will get out of sync and crash. (revdb.py can usually recover from such crashes and let you continue.)

More notes:

  • When revdb.py is busy moving in time, it prints the progress, for example as (1500000...). If you messed up, or simply are not interested in it continuing searching after a while, you can safely press Ctrl-C to have it stop and jump back to the timestamp it was previously at. This is particularly important with watchpoints, because they make running a lot slower. (You should anyway delete watchpoints when their role has been fulfilled, but in the future we might cache the watchpoint results so that they are only evaluated the first time we go over each timestamp.)
  • Setting a watchpoint or printing a $NUM in the past requires a rescan of the log file from the time of creation of that object (once). If $NUM is an object created very early in the process, you will have to wait (or use Ctrl-C).
  • When tracking a complex bug, it is recommended to write down the timeline on a piece of paper (or separate file). Make sure you write the timestamp for every event you record, and keep the log ordered by timestamp. Write down which $NUM corresponds to the relevant objects. All the timestamps that you write down are still valid if you leave and restart revdb.py. The $NUM are not, though. (This might be changed in the future. For now it should be easy to rebuild them manually by using go TIMESTAMP and repeating the print commands.)

Contact information

IRC: #pypy on irc.freenode.net

Mailing list: pypy-dev@python.org

You can report issues in the issue tracker of RevDB.

RevDB is made by Armin Rigo, but thanks go to the rest of the PyPy team as well.