Comparisons
In this section we compare typedload to other similar libraries.
In general, the advantages of typedload over competing libraries are:
- Easy to use
- Very fast when Unions are involved
- Works with existing codebase and uses standard types. No inheritance or decorators
- Easy to extend, even with objects from 3rd party libraries
- Stable API, breaking changes only happen on major releases (it has happened once since 2018 and most users didn't notice)
- Mypy and similar work without plugins
- Can use and convert camelCase and snake_case
- Functional approach
- Pure Python, no compiling
- Very small, it's fast for automated tests to download, extract and install compared to huge binary libraries
It works with existing codebase
Most libraries require your classes to extend or use decorators from the library itself.
This means that types from other libraries or non supported stdlib classes can never be used.
It also means that mypy will just work out of the box, rather than requiring plugins.
Instead, typedload works fine with the type annotations from the typing
module and will work without requiring any changes to the datatypes.
It is easy to extend
Since there can be situations that are highly domain specific, typedload allows to extend its functionality to support more types or replace the existing code to handle special cases.
Support of Union
Other libraries tend to either be very slow or just give completely wrong results when Union are involved. Typedload works without having to manually do annoying annotations.
Functional approach
You can load a list[YourType]
, without having to create a loader object or a useless container object.
apischema
Found here
It's the only viable alternative to typedload that I've encountered.
- Settings are global, a submodule changing a setting will affect the entire application
- Type checks are disabled by default
- It reuses the same objects in the output, so changing the data might result in subtle bugs if the input data is used again
- No native support for attrs (but can be manually added by the user)
pydantic
Found here
- Complicated API
- The author calls you a liar if your pure python library is faster
- Breaks API all the time, between minor releases. (43 times in 2 major versions so far)
- They hate benchmarks that show it's slow. So they removed them altogether
- It needs a mypy plugin, and for some kinds of classes it does no static checks whatsoever.
- Is now VC funded, so eventually some draconian monetizing plan might appear.
Version 1
- One of the slowest libraries that exist in this space
int | float
might decide to cast afloat
toint
, or anint
tofloat
Version 2
- Despite the rewrite in rust, and taking inspiration from typedload's autotagging of unions somehow manages to be slower than pure python to load unions.
- Took them several years to make a version 2 where types on BaseModel finally mean the same thing that they mean in the rest of python
- Took them several years to implement unions that don't cast types at random
jsons
Found here
- Type safety is not a goal of this project
- It is buggy:
- This returns an int
jsons.load(1.1, Union[int, float])
- This returns a string
jsons.load(1, Union[str, int])
- This raises an exception
jsons.load(1.0, int | float)
- This returns an int
- It is incredibly slow (40x slower in certain cases) For this reason it has been removed from the benchmarks.
- Does not support
Literal
- Can't load iterables as lists
- Exceptions do not have information to find the incorrect data
Quick test
# Needed because jsons can't load directly from range()
data = [i for i in range(3000000)]
# This took 2.5s with jsons and 200ms with typedload
load(data, list[int])
# This took 20s with jsons and 500ms with typedload
# And it converted all the ints to float!
load(data, list[Union[float,int]])
dataclasses-json
Found here
It is completely useless for type safety and very slow. I can't understand why it has users.
- It is incredibly slow (20x slower in certain cases) For this reason it has been removed from the benchmarks.
- It doesn't enforce type safety (it will be happy to load whatever inside an int field)
- Requires to decorate all the classes
- It is not extensible
- Has dependencies (marshmallow, marshmallow-enum, typing-inspect)
- Very complicated way for doing lists
Quick test
# Just to show how unreliable this is
@dataclass_json
@dataclass
class Data:
data: list[int]
Data.from_dict({'data': ["4", None, ..., ('qwe',), 1.1]})
# This returns:
# Data(data=['4', None, Ellipsis, ('qwe',), 1.1])
# despite data is supposed to be a list of int
msgspec
- Very fast, but unions don't work
- The author will send you a PR to add his project to the benchmarks but will refuse to add your project to his benchmarks when you do the same, saying that your project is not popular enough (despite it having many more downloads)
In theory unions do work, but you need to refactor your entire project around using msgspec's peculiar idea of unions to use them, and even then they are very limited in scope, compared to what other projects support and python users normally use.
- Implemented in C, won't run on PyPy
- Supports tagged Unions partially only when inheriting from its Struct type Mypy will not typecheck those classes. To use unions you must give up static typechecking.
- Doesn't support unions between regular dataclass/NamedTuple/Attrs/TypedDict
- Doesn't support untagged Unions
- Doesn't support multiple tags (e.g.
tag=Literal[1, 2]
) - Extended using a single function that must handle all cases
- Can't replace type handlers