The windows failure is known and unrelated: #109739

I don't think this change is worth making. It seems mostly like code churn: there's no actual problem being solved.

Is there a difference if the functools import is deferred to within the vendored _recursive_repr function (& we remove the top-level)?

A

@AA-Turner it makes no difference, because inspect imports 80% of all stdlib, including functools :)

inspect is only used in two places:

• cpython/Lib/dataclasses.py

  Line 970 in f2eaa92

  cls_annotations = inspect.get_annotations(cls)

• cpython/Lib/dataclasses.py

  Line 1139 in f2eaa92

  text_sig = str(inspect.signature(cls)).replace(' -> None', '')

But, it still does not improve the import time:

» pyperf timeit 'import dataclasses'
.....................
Mean +- std dev: 105 ns +- 1 ns

I also tried to defer inspect, but it is still the same :(

...it makes no difference, because inspect imports 80% of all stdlib, including functools :)

inspect is only used in two places:

• cpython/Lib/dataclasses.py

  Line 970 in f2eaa92

  cls_annotations = inspect.get_annotations(cls)

• cpython/Lib/dataclasses.py

  Line 1139 in f2eaa92

  text_sig = str(inspect.signature(cls)).replace(' -> None', '')

But, it still does not improve the import time:

» pyperf timeit 'import dataclasses'
.....................
Mean +- std dev: 105 ns +- 1 ns

I also tried to defer inspect, but it is still the same :(

@sobolevn
Are these pyperf timings actually outputting accurate data when it comes to imports? The numbers seem surprisingly low to me.

I get similar numbers (scale-wise) from pyperf but they don't match with other timings I've done.

pyperf timeit "import dataclasses" ¹

.....................
Mean +- std dev: 84.6 ns +- 4.1 ns

python -Ximporttime -c "import dataclasses" (truncated to only show dataclasses):

import time: self [us] | cumulative | imported package
...
import time:       961 |      18112 | dataclasses

Which lines up roughly with the timings I get from hyperfine comparing launching python with no imports to importing dataclasses.

hyperfine -w10 -r100 -N 'python -c "pass"' 'python -c "import dataclasses"'

Benchmark 1: python -c "pass"
  Time (mean ± σ):      11.0 ms ±   0.6 ms    [User: 8.4 ms, System: 2.4 ms]
  Range (min … max):    10.4 ms …  13.0 ms    100 runs
 
Benchmark 2: python -c "import dataclasses"
  Time (mean ± σ):      28.5 ms ±   1.0 ms    [User: 23.8 ms, System: 4.5 ms]
  Range (min … max):    27.0 ms …  32.0 ms    100 runs
 
Summary
  'python -c "pass"' ran
    2.59 ± 0.17 times faster than 'python -c "import dataclasses"'

From previous experiments mentioned in #109653 (comment) I would expect deferring inspect to remove close to half of the import time. However, it's not actually very useful to do so as any use of @dataclass will end up importing inspect in order to get the annotations, so practically you wouldn't actually see any benefit from delaying the import unless none of the modules you use make any dataclasses.

Slightly more directly related, and also mentioned in #109653 (comment) the definition of _recursive_repr you have in dataclasses makes use of functools.wraps which uses functools.partial which in turn uses reprlib.recursive_repr on its own __repr__ so perhaps this change still makes sense from that angle?

A change in the other direction (per reprlib's own implementation):

diff --git a/Lib/dataclasses.py b/Lib/dataclasses.py
index 84f8d68ce0..8579f6d605 100644
--- a/Lib/dataclasses.py
+++ b/Lib/dataclasses.py
@@ -4,7 +4,6 @@
 import types
 import inspect
 import keyword
-import functools
 import itertools
 import abc
 import _thread
@@ -252,7 +251,6 @@ def _recursive_repr(user_function):
     # call.
     repr_running = set()

-    @functools.wraps(user_function)
     def wrapper(self):
         key = id(self), _thread.get_ident()
         if key in repr_running:
@@ -263,6 +261,15 @@ def wrapper(self):
         finally:
             repr_running.discard(key)
         return result
+
+    wrapper.__module__ = getattr(user_function, '__module__')
+    wrapper.__doc__ = getattr(user_function, '__doc__')
+    wrapper.__name__ = getattr(user_function, '__name__')
+    wrapper.__qualname__ = getattr(user_function, '__qualname__')
+    wrapper.__annotations__ = getattr(user_function, '__annotations__', {})
+    wrapper.__type_params__ = getattr(user_function, '__type_params__', ())
+    wrapper.__wrapped__ = user_function
+
     return wrapper

 class InitVar:

A

@AA-Turner we still have _thread import (1 _thread vs 1 reprlib) and quite a lot of boilerplate code that can go out of sync with the reprlib.recursive_repr (it happened before #109819) for no real reason.

As long as there's not a performance regression due to importing reprlib, I'm okay with this. I have not checked the performance.

You're already importing reprlib indirectly via functools so this shouldn't change import time if that's a concern (functools will still be imported via inspect so you won't see an improvement by removing that direct import).

Agreed that reprlib is already imported (checked with 3.11).