Repository navigation
Description
Documentation
Hey.
The documentation of LogRecord describes the exc_info attribute as:
exc_info (tuple[type[BaseException], BaseException, types.TracebackType] | None) – An exception tuple with the current exception information, as returned by sys.exc_info(), or None if no exception information is available.
However, records created through the normal logger methods can contain values that don’t match this type:
import logging
class H(logging.Handler):
def emit(self, record):
print(repr(record.exc_info))
logger = logging.getLogger()
logger.addHandler(H())
logger.propagate = False
logger.error("b", exc_info=True) # outside of an except-block
logger.error("c", exc_info=ValueError()) # exception that was never raisedwhich gives as of CPython 3.14.10:
(None, None, None)
(<class 'ValueError'>, ValueError(), None)
A very recent commit (938144e) already fixed the case of a false truth value, which as per “If exc_info does not evaluate as false, …” is IMO allowed.
Now:
- Outside an
except-block it would be(None, None, None). - If an exception instance is passed, the tuple is built from its traceback, which is None for exceptions that were never raised, while the documented type requires types.TracebackType.
I think typeshed already documents it properly:
https://github.com/python/typeshed/blob/80f7beabfbb71a634cc3dbca8e418a5da790142c/stdlib/logging/__init__.pyi#L63
So far I'd say this is a documentation issue.
Now I can create valid LogRecord-objects that don’t go through _log() and its normalisation:
>>> import logging
>>> l = logging.getLogger()
>>> r = logging.makeLogRecord({"exc_info": True})
>>> r.exc_info
True
>>> r = l.makeRecord("z", 40, "f", 1, "m", None, exc_info=False)
>>> r.exc_info
False
>>>
For makeLogRecord and makeRecord… it’s IMO not really clear what values these allow (de jure) for exc_info.
If they’d follow the spirit of “If exc_info does not evaluate as false, …” then the the above type description would be even more wrong, as then exc_info could be any object.
This would then IMO also make the types on typeshed wrong, cause while logging.Formatter.format() itself would call formatException() only if exc_info has a true truth value:
cpython/Lib/logging/__init__.py
Lines 716 to 720 in ecefd40
| if record.exc_info: | |
| # Cache the traceback text to avoid converting it multiple times | |
| # (it's constant anyway) | |
| if not record.exc_text: | |
| record.exc_text = self.formatException(record.exc_info) |
that may be still any object, there's nothing for the false case and both format() and formatException() are public, so any subclass may call formatException() without the if record.exc_info:.
Last but not least:
formatException() is more or less documented to an even more restrictive type:
Formats the specified exception information (a standard exception tuple as returned by sys.exc_info()) as a string.
So not sure,... at least some documentation seems to be not exact... but maybe there also needs to be code changes if the normalisation of exc_info should happen not only via LogRecord objects created through _log()... and if not, typeshed might need some updates to account for values that may be given.
Cheers,
Chris
Metadata
Metadata
Assignees
Labels
Projects
- StatusShow more project fieldsNo status
- StatusShow more project fieldsTodo
Activity
calestyo commentedon Nov 13, 2026
Perhaps to clarify on the potential effect on typeshed, which currently has:
But when created via
makeLogRecordormakeRecord, no normalisation seems to happen andexc_infomight beobject.calestyo commentedon Nov 13, 2026
I’ve just noted that the same problem exists in principle for
LogRecord.stack_info, too.LogRecordeven defaults to astack_infoof None.With
logging.Formatter.format()itself that value would just never be passed on toformatStack():cpython/Lib/logging/__init__.py
Lines 725 to 728 in ecefd40
But nothing seems to make this behaviour mandatory for any derived versions of
format().I think this again affects typeshed, which allows only
str:https://github.com/python/typeshed/blob/80f7beabfbb71a634cc3dbca8e418a5da790142c/stdlib/logging/__init__.pyi#L290
anbv29 commentedon Nov 15, 2026
Hey, could you assign this to me if it’s still up for grabs? I’d start with a docs patch for the two cases from the repro. One is logging outside an exception handler and the other is passing an exception that was never raised. Does it make sense to keep
makeLogRecord()andmakeRecord()out of scope for this patch?