Castrated traceback in sys.exc_info()

[Pakal]

Hello

I've just realized recently that sys.exc_info() didn't return a full
traceback for exception concerned : it actually only contains the
frame below the point of exception catching.

That's very annoying to me, because I planned to log such tracebacks
with logging.critical(*****, exc_info=True), and these partial
tracebacks, like the one below, are clearly unsufficient to determine
where the problem comes from. A whole traceback, from program entry
point to exception raising point, would be much better.

2010-03-17 09:28:59,184 - pims - CRITICAL - This is just a test to
ensure critical emails are properly sent
Traceback (most recent call last):
<< HERE, lots of frames missing >>
File "test_common.py", line 34, in test_email_sending
os.open("qsdsdqsdsdqsd", "r")
TypeError: an integer is required

Is there any workaround for this ? I've thought about a custom logging
formatter, which would take both the exc_info traceback AND its own
full backtrace, and to connect them together on their relevant part,
but it's awkward and error prone... why can't we just have all the
precious traceback info under the hand there (an additional attribute
might have pointed the precise frame in which the exception was
caught).

Tanks for the attention,
Regards,
Pascal

 

[Michael Ricordeau]

Hi,

to log tracebacks, you can probably try traceback module.

I use it like this :

import traceback
..... # your code

for line in traceback.format_exc().splitlines():
log.trace(line)



Le Wed, 17 Mar 2010 03:42:44 -0700 (PDT),

 

[Vinay Sajip]

Hello

I've just realized recently that sys.exc_info() didn't return a full
traceback for exception concerned : it actually only contains the
frame below the point of exception catching.

That's very annoying to me, because I planned to log such tracebacks
withlogging.critical(*****, exc_info=True), and these partial
tracebacks, like the one below, are clearly unsufficient to determine
where the problem comes from. A whole traceback, from program entry
point to exception raising point, would be much better.

2010-03-17 09:28:59,184 - pims - CRITICAL - This is just a test to
ensure critical emails are properly sent
Traceback (most recent call last):
<< HERE, lots of frames missing >>
  File "test_common.py", line 34, in test_email_sending
    os.open("qsdsdqsdsdqsd", "r")
TypeError: an integer is required

Is there any workaround for this ? I've thought about a customlogging
formatter, which would take both the exc_info traceback AND its own
full backtrace, and to connect them together on their relevant part,
but it's awkward and error prone... why can't we just have all the
precious traceback info under the hand there (an additional attribute
might have pointed the precise frame in which the exception was
caught).

Tanks for the attention,
Regards,
Pascal

Do you have a short script which demonstrates the problem? Some more
context is needed. For example, if you have multiple threads of
execution, the traceback will only go up to the top-level function in
the thread.

Regards,

Vinay Sajip

 

[Gabriel Genellina]

En Wed, 17 Mar 2010 09:42:06 -0300, Pascal Chambon

traceback functions indeed allow the manipulation of exception
tracebacks,
but the root problem is that anyway, since that traceback is incomplete,
your "traceback.format_exc().splitlines()" will only provide frames for
callee (downward) functions, not caller (upward) ones, starting from the
exception catching frame.

Either I don't understand what you mean, or I can't reproduce it:

<code>
import logging

def a(): return b()
def b(): return c()
def c(): return d()
def d(): raise ValueError

def main():
logging.basicConfig(level=logging.DEBUG)
try: a()
except: logging.exception("An error")

main()
</code>

Output:

D:\temp>python test_logging.py
ERROR:root:An error
Traceback (most recent call last):
File "test_logging.py", line 10, in main
try: a()
File "test_logging.py", line 3, in a
def a(): return b()
File "test_logging.py", line 4, in b
def b(): return c()
File "test_logging.py", line 5, in c
def c(): return d()
File "test_logging.py", line 6, in d
def d(): raise ValueError
ValueError

 

[Pascal Chambon]

Gabriel Genellina a écrit :

En Wed, 17 Mar 2010 09:42:06 -0300, Pascal Chambon


Either I don't understand what you mean, or I can't reproduce it:

Allright, here is more concretely the problem :

<code>
import logging

def a(): return b()
def b(): return c()
def c():
try:
return d()
except:
logging.exception("An error")

def d(): raise ValueError

def main():
logging.basicConfig(level=logging.DEBUG)
a()

main()
</code>

OUTPUT:ERROR:root:An error
Traceback (most recent call last):
File "C:/Users/Pakal/Desktop/aaa.py", line 7, in c
return d()
File "C:/Users/Pakal/Desktop/aaa.py", line 11, in d
def d(): raise ValueError
ValueError

As you see, the traceback only starts from function c, which handles the
exception.
It doesn't show main(), a() and b(), which might however be (and are, in
my case) critical to diagnose the severity of the problem (since many
different paths would lead to calling c()).

So the question is : is that possible to enforce, by a way or another,
the retrieval of the FULL traceback at exception raising point, instead
of that incomplete one ?

Thank you for your help,
regards,

Pascal

 

[Gabriel Genellina]

En Mon, 22 Mar 2010 15:20:39 -0300, Pascal Chambon

Allright, here is more concretely the problem :

ERROR:root:An error
Traceback (most recent call last):
File "C:/Users/Pakal/Desktop/aaa.py", line 7, in c
return d()
File "C:/Users/Pakal/Desktop/aaa.py", line 11, in d
def d(): raise ValueError
ValueError

As you see, the traceback only starts from function c, which handles the
exception.
It doesn't show main(), a() and b(), which might however be (and are, in
my case) critical to diagnose the severity of the problem (since many
different paths would lead to calling c()).

So the question is : is that possible to enforce, by a way or another,
the retrieval of the FULL traceback at exception raising point, instead
of that incomplete one ?

Thanks for bringing this topic! I learned a lot trying to understand what
happens.

The exception traceback (what sys.exc_info()[2] returns) is *not* a
complete stack trace. The sys module documentation is wrong [1] when it
says "...encapsulates the call stack at the point where the exception
originally occurred."

The Language Reference is more clear [2]: "Traceback objects represent a
stack trace of an exception. A traceback object is created when an
exception occurs. When the search for an exception handler unwinds the
execution stack, at each unwound level a traceback object is inserted in
front of the current traceback. When an exception handler is entered, the
stack trace is made available to the program."

That is, a traceback holds only the *forward* part of the stack: the
frames already exited when looking for an exception handler. Frames going
from the program starting point up to the current execution point are
*not* included.

Conceptually, it's like having two lists: stack and traceback. The
complete stack trace is always stack+traceback. At each step (when
unwinding the stack, looking for a frame able to handle the current
exception) an item is popped from the top of the stack (last item) and
inserted at the head of the traceback.

The traceback holds the "forward" path (from the current execution point,
to the frame where the exception was actually raised). It's a linked list,
its tb_next attribute holds a reference to the next item; None marks the
last one.

The "back" path (going from the current execution point to its caller and
all the way to the program entry point) is a linked list of frames; the
f_back attribute points to the previous one, or None.

In order to show a complete stack trace, one should combine both. The
traceback module contains several useful functions: extract_stack() +
extract_tb() are a starting point. The simplest way I could find to make
the logging module report a complete stack is to monkey patch
logging.Formatter.formatException so it uses format_exception() and
format_stack() combined (in fact it is simpler than the current
implementation using a StringIO object):

<code>
import logging
import traceback

def formatException(self, ei):
"""
Format and return the specified exception information as a string.

This implementation builds the complete stack trace, combining
traceback.format_exception and traceback.format_stack.
"""
lines = traceback.format_exception(*ei)
if ei[2]:
lines[1:1] = traceback.format_stack(ei[2].tb_frame.f_back)
return ''.join(lines)

# monkey patch the logging module
logging.Formatter.formatException = formatException

def a(): return b()
def b(): return c()
def c():
try:
return d()
except:
logging.exception("An error")
raise
def d(): raise ValueError

def main():
a()

main()
</code>

Output:

ERROR:root:An error
Traceback (most recent call last):
File "test_logging.py", line 32, in <module>
main()
File "test_logging.py", line 30, in main
a()
File "test_logging.py", line 19, in a
def a(): return b()
File "test_logging.py", line 20, in b
def b(): return c()
File "test_logging.py", line 23, in c
return d()
File "test_logging.py", line 27, in d
def d(): raise ValueError
ValueError

Traceback (most recent call last):
File "test_logging.py", line 32, in <module>
main()
File "test_logging.py", line 30, in main
a()
File "test_logging.py", line 19, in a
def a(): return b()
File "test_logging.py", line 20, in b
def b(): return c()
File "test_logging.py", line 23, in c
return d()
File "test_logging.py", line 27, in d
def d(): raise ValueError
ValueError

Note that both tracebacks are identical: the first comes from the patched
logging module, the second is the standard Python one.

[1] http://docs.python.org/library/sys.html#sys.exc_info
[2]
http://docs.python.org/reference/datamodel.html#the-standard-type-hierarchy

 

[Pascal Chambon]

Gabriel Genellina a écrit :

En Mon, 22 Mar 2010 15:20:39 -0300, Pascal Chambon

Allright, here is more concretely the problem :

ERROR:root:An error
Traceback (most recent call last):
File "C:/Users/Pakal/Desktop/aaa.py", line 7, in c
return d()
File "C:/Users/Pakal/Desktop/aaa.py", line 11, in d
def d(): raise ValueError
ValueError

As you see, the traceback only starts from function c, which handles
the exception.
It doesn't show main(), a() and b(), which might however be (and are,
in my case) critical to diagnose the severity of the problem (since
many different paths would lead to calling c()).

So the question is : is that possible to enforce, by a way or
another, the retrieval of the FULL traceback at exception raising
point, instead of that incomplete one ?

Thanks for bringing this topic! I learned a lot trying to understand
what happens.

The exception traceback (what sys.exc_info()[2] returns) is *not* a
complete stack trace. The sys module documentation is wrong [1] when
it says "...encapsulates the call stack at the point where the
exception originally occurred."

The Language Reference is more clear [2]: "Traceback objects represent
a stack trace of an exception. A traceback object is created when an
exception occurs. When the search for an exception handler unwinds the
execution stack, at each unwound level a traceback object is inserted
in front of the current traceback. When an exception handler is
entered, the stack trace is made available to the program."

That is, a traceback holds only the *forward* part of the stack: the
frames already exited when looking for an exception handler. Frames
going from the program starting point up to the current execution
point are *not* included.

Conceptually, it's like having two lists: stack and traceback. The
complete stack trace is always stack+traceback. At each step (when
unwinding the stack, looking for a frame able to handle the current
exception) an item is popped from the top of the stack (last item) and
inserted at the head of the traceback.

The traceback holds the "forward" path (from the current execution
point, to the frame where the exception was actually raised). It's a
linked list, its tb_next attribute holds a reference to the next item;
None marks the last one.

The "back" path (going from the current execution point to its caller
and all the way to the program entry point) is a linked list of
frames; the f_back attribute points to the previous one, or None.

In order to show a complete stack trace, one should combine both. The
traceback module contains several useful functions: extract_stack() +
extract_tb() are a starting point. The simplest way I could find to
make the logging module report a complete stack is to monkey patch
logging.Formatter.formatException so it uses format_exception() and
format_stack() combined (in fact it is simpler than the current
implementation using a StringIO object):

Good point, there is clearly a distinction between "stack trace" and
"exception traceback" that I didn't know (actually, it seems no one
makes it in computer literature).

Good catch, Gabriel.

There should be no need to monkey-patch the logging module - it's
better if I include the change in the module itself. The only
remaining question is that of backward compatibility, but I can do
this for Python 2.7/3.2 only so that won't be an issue. It's probably
a good idea to log an issue on the bug tracker, though, so we have
some history for the change - do you want to do that, or shall I?

Regards,

Vinay Sajip

Well having it fixed in logging would be great, but that kind of
information is good to have in other circumstances, so shouldn't we
rather advocate the availability of this "stack trace part" in exc_info
too ?
This way, people like me who consider frames as black magic wouldn't
need to meet complex stuffs as
"traceback.format_stack(ei[2].tb_frame.f_back"

Should I open an issue for this evolution of exceptiuon handling, or
should we content ourselves of this "hacking" of frame stck ?

Regards,
Pascal

 

[Vinay Sajip]

Should I open an issue for this evolution of exceptiuon handling, or
should we content ourselves of this "hacking" of frame stck ?

Possibly worth raising an issue (not logging-related), but perhaps
it's worth seeing if this has come up before creating the issue.

Regards,

Vinay Sajip

 

[Gabriel Genellina]

Possibly worth raising an issue (not logging-related), but perhaps
it's worth seeing if this has come up before creating the issue.

There is, from 2006: http://bugs.python.org/issue1553375
I'll add a patch+tests to this issue.