jcrist / msgspec

A fast serialization and validation library, with builtin support for JSON, MessagePack, YAML, and TOML
https://jcristharif.com/msgspec/
BSD 3-Clause "New" or "Revised" License
2.43k stars 75 forks source link

Encoding exceptions - segfault #727

Closed regnarg closed 1 month ago

regnarg commented 2 months ago

Question

I tried to make serializable exceptions using:

import msgspec
class TaggedException(msgspec.Struct, Exception, tag=True): pass

However, this reliably segfaults Python (CPython 3.12.5, msgspec 0.18.6).

I presume there is some kind of incompatibility between the C-level layout of Struct objects and Exception objects? Are there any good alternatives for encoding exception objects? (which seems super usefil when implementing APIs) Maybe this cannot be fixed. But maybe it should at least throw an error instead of crashing the interpreter?

provinzkraut commented 2 months ago

I usually do something like this:


class EncodableException(Exception):
  def to_dict(self) -> dict[str, Any]:
   ...

def enc_hook(value: Any) -> Any:
  if isinstance(value, EncodableException):
    return value.to_dict()
  raise TypeError()

and then use this enc_hook whenever I'm encoding stuff that might contain one of those exceptions.

jcrist commented 1 month ago

Thanks for opening this. In prior versions msgspec had a check preventing mixing Struct classes with other builtin types - this check was broken on Python 3.12 but is now fixed in #745. The above example now errors nicely rather than segfaulting. If you want to encode exceptions I recommend following a pattern like the one @provinzkraut outlined above rather than defining an Exception type that is also a Struct.