Repository navigation
Structs with nested array_like Structs asdict #830
Description
Question
Hey folks, I went over #594 because of the same assumptions.
I was expecting asdict to work on nested Structs, at least as an option.
This is where I'm struggling:
import msgspec
class Dog(msgspec.Struct, array_like=True):
name: str
age: int
class Owner(msgspec.Struct):
name: str
dogs: list[Dog]
owner = Owner('A dog owner', [Dog('Doggie', 2), Dog('Max', 1)])
msgspec.structs.asdict(owner)
>> {'name': 'A dog owner', 'dogs': [Dog(name='Doggie', age=2), Dog(name='Max', age=1)]}
msgspec.to_builtins(owner)
>> {'name': 'A dog owner', 'dogs': [['Doggie', 2], ['Max', 1]]}
I need to come up with the following:
>>> {'name': 'A dog owner', 'dogs': [{'name': 'Doggie', 'age': 2}, {'name': 'Max', 'age': 1}]}
I tried working around the issue with the following, but most of the performance gains vanish.
def asdict(obj):
if isinstance(obj, msgspec.Struct):
return asdict(msgspec.structs.asdict(obj))
if isinstance(obj, dict):
return {key: asdict(value) for key, value in obj.items()}
if isinstance(obj, list):
return [asdict(item) for item in obj]
return obj
Is there any plan to support the nested structs this way in asdict in the short term? or can you guys think of a good workaround? Thank you
Activity
Expand list of differences between asdict/astuple and to_builtins
Siyet commentedon May 26, 2026
Apologies for the long silence on this.
The behavior is intentional:
msgspec.structs.asdictis shallow by design (#299 and #297). The two intended paths are:msgspec.structs.asdict/astuple- one-to-one conversion of a single struct instance, no recursion, no encoder settings applied. Modeled afterdataclasses.asdict/dataclasses.astuplebut, unlike those, doesn't descend into nested values.msgspec.to_builtins- the encoding pipeline, recurses into nested structs and applies all encoder settings, includingarray_like.Your case sits in between: you want recursion (like
to_builtins) but notarray_like(likeasdict). msgspec doesn't expose a third mode for that. Witharray_like=TrueonDog,to_builtins(owner)will always emit lists for the dogs - that's the trade-offarray_likeopts you into across the entire encoder surface.The new docs section in #1025 spells the split out explicitly: Converting to and from Builtin Types. Hopefully that makes the boundary easier to find next time.
For the dict-with-nested-dicts output, the cleanest workaround is a small recursive helper that walks
__struct_fields__:For your example:
This is pure Python and noticeably slower than the C path (in a quick local bench it's ~7x slower than
to_builtins(owner)), but for one-off conversions outside hot loops it's usually fine.Closing as not planned since adding an
ignore_array_like(or similar) flag conflicts with the deliberate split betweenasdictandto_builtins. If you have a use case where the recursive helper isn't sufficient, happy to reopen the discussion.