Omitting defaults does not work with array-like structs

[chuckwondo]

Description

When setting both omit_defaults and array_like to True, defaults are not omitted.

For example, without array_like=True:

class Position(
    msgspec.Struct,
    frozen=True,
    forbid_unknown_fields=True,
    omit_defaults=True,
):
    longitude: float
    latitude: float
    altitude: float = 0.0

Performing a roundtrip works as expected:

>>> pos = msgspec.json.Decoder(Position).decode('{ "longitude": 1.0, "latitude": 2.0 }')
>>> pos
Position(longitude=1.0, latitude=2.0, altitude=0.0)
>>> msgspec.json.encode(pos)
b'{"longitude":1.0,"latitude":2.0}'

However, adding array_like=True to the Position definition above causes rountripping to fail:

>>> pos = msgspec.json.Decoder(Position).decode('[1.0,2.0]')
>>> pos
Position(longitude=1.0, latitude=2.0, altitude=0.0)
>>> msgspec.json.encode(pos)
b'[1.0,2.0,0.0]'

[ipseitas]

Hi there!

Just noticed this yesterday as well.
From what I can see ,this doesn't only affect the json encoding.
This is an example of MsgPack acting the same way:

from msgspec.msgpack import decode, encode
from msgspec.structs import Struct

class User(Struct, tag=1, array_like=True):
    username: str
    is_admin: bool = False

class Comment(Struct, tag=2, array_like=True):
    user: User
    content: str
    likes: int = 0
    is_highlighted: bool = False

username = "ipseitas"
content = "Lovin' this lib, please add donations!"

comment = Comment(User(username), content)
commentlike = (2, (1, username), content)

>>> encode(comment)
... b"\x95\x02\x93\x01\xa8ipseitas\xc2\xd9&Lovin' this lib, please add donations!\x00\xc2"

>>> encode(commentlike)
... b"\x93\x02\x92\x01\xa8ipseitas\xd9&Lovin' this lib, please add donations!"

>>> decode(encode(commentlike), type=Comment) == comment
... True

Implementing fits the theme of being efficient that your package shines at.
If this helps you deliver, please add donation #542.
I'll gladly transfer some of the time your lib saved me and am sure others would as well :)

[LukasKrocek]

It is not possible to omit in some cases. Imagine:

ENCODER = msgspec.msgpack.Encoder()

class A(msgspec.Struct):
    a: bool = False
    b: bool = False

ENCODER.encode(A(b=True)) # would be [True]

[ipseitas]

It's possible if we assume a specific behavior.
For example, only strip from the end of the array.
That way the fields that are more likely to be modified could be put to the front.

This Struct:

class User(Struct, array_like=True):
    username: str
    profile_pic_url: str | None = None
    is_admin: bool = False

Could be:

["robert"] # has no pic, not an admin
["jon", None, True] # is an admin, so we need to keep the default pic value anyway

[ipseitas]

Another behavior could be a flag that allows the gap to be filled with None.
That way, the encoder would save space on non-nullable fields, and the decoder would use the default value.

So that this message:

class TransportMethod(StrEnum):
    One = "one"
    Two = "two"
    AnotherOne = "another_one"

class Package(Struct, array_like=True, lax_array_defaults=True):
    content: str
    to_dispatcher: TransportMethod = TransportMethod.One
    from_dispatcher: TransportMethod = TransportMethod.One

Could encode to these:

["content"] # Uses default transport methods
["content", None, "two"] # Only one transport is modified

But then this message:

class Package(Struct, array_like=True, lax_array_defaults=True):
    content: str
    to_dispatcher: TransportMethod | None = TransportMethod.One
    from_dispatcher: TransportMethod = TransportMethod.One

Would be encoded as:

["content"] # Uses default transport methods
["content", "one", "two"] # Only one transport is modified, but setting the first one to None would change its meaning