use default allocator

[adriangb]

No description provided.

[codspeed-hq]

CodSpeed Performance Report

Merging #662 default-allocator (ac54212) will degrade performances by 15.0%.

Summary

🔥 0 improvements
❌ 80 regressions
✅ 45 untouched benchmarks

🆕 0 new benchmarks
⁉️ 0 dropped benchmarks

⚠️ Please fix the performance issues or acknowledge them on CodSpeed.

Benchmarks breakdown

	Benchmark	main	default-allocator	Change
❌	test_complete_core_lax	1.3 ms	1.5 ms	-14.35%
❌	test_complete_core_strict	1.4 ms	1.5 ms	-10.76%
❌	test_complete_core_error	7.9 ms	9.8 ms	-24.38%
❌	test_complete_core_isinstance	7.7 ms	9.8 ms	-26.25%
❌	test_complete_core_json	2.8 ms	3.3 ms	-19.80%
❌	test_build_schema	3.9 ms	4.6 ms	-15.67%
❌	test_core_python_fs	345 µs	409.3 µs	-18.65%
❌	test_core_python	512.8 µs	572.4 µs	-11.64%
❌	test_core_json_fs	667.8 µs	861.1 µs	-28.95%
❌	test_bool_core	49.4 µs	54.5 µs	-10.18%
❌	test_small_class_core_dict	26.1 µs	32.5 µs	-24.53%
❌	test_small_class_core_model	45.2 µs	57.3 µs	-26.83%
❌	test_core_string_lax_wrong	35.2 µs	42.9 µs	-21.89%
❌	test_core_string_strict_wrong	34.9 µs	42.5 µs	-21.89%
❌	test_core_string_strict_wrong_str_e	64.4 µs	79.1 µs	-22.68%
❌	test_definition_model_core	1.9 ms	2.1 ms	-15.58%
❌	test_list_of_dict_models_core	486.4 µs	556 µs	-14.32%
❌	test_list_of_ints_core_json	5.5 ms	6.6 ms	-20.74%
❌	test_set_of_ints_core_duplicates	2.4 ms	2.8 ms	-19.18%
❌	test_set_of_ints_core_json	6.8 ms	7.7 ms	-13.29%
❌	test_set_of_ints_core_json_duplicates	4.9 ms	5.8 ms	-17.11%
❌	test_frozenset_of_ints_core	1.3 ms	1.5 ms	-11.57%
❌	test_frozenset_of_ints_duplicates_core	993.9 µs	1,182.8 µs	-19.00%
❌	test_dict_of_any_core	4.2 ms	4.9 ms	-16.93%
❌	test_dict_of_ints_core_json	13.9 ms	16.3 ms	-17.33%
❌	test_many_models_core_dict	5 ms	6.1 ms	-22.28%
❌	test_many_models_core_model	14.3 ms	16.1 ms	-12.31%
❌	test_list_of_nullable_core	637.4 µs	809.4 µs	-26.98%
❌	test_core_python	39.8 µs	49.2 µs	-23.63%
❌	test_model_core_json	52.3 µs	62.7 µs	-19.87%
❌	test_core_str	25.4 µs	31.1 µs	-22.19%
❌	test_core_future	41.4 µs	48.5 µs	-17.23%
❌	test_core_future_str	26.1 µs	31.7 µs	-21.42%
❌	test_date_from_str	20.7 µs	27.7 µs	-33.81%
❌	test_date_from_datetime	33 µs	41 µs	-24.23%
❌	test_date_from_datetime_str	23.8 µs	30.5 µs	-28.09%
❌	test_core_future	24.3 µs	30.3 µs	-24.92%
❌	test_core_future_str	20.9 µs	27.1 µs	-29.25%
❌	test_strict_union_error_core	41.2 µs	46.6 µs	-13.06%
❌	test_raise_error_value_error	54.8 µs	62.6 µs	-14.13%
❌	test_raise_error_custom	56.4 µs	63.6 µs	-12.70%
❌	test_positional_tuple	26.3 µs	31.4 µs	-19.53%
❌	test_variable_tuple	26.6 µs	31.6 µs	-18.87%
❌	test_tuple_many_variable	29.4 µs	35.5 µs	-21.03%
❌	test_tuple_many_positional	30 µs	37.2 µs	-23.98%
❌	test_arguments	36.8 µs	44.7 µs	-21.45%
❌	test_with_default	36.4 µs	42.6 µs	-17.00%
❌	test_chain_list	37.1 µs	41.9 µs	-13.12%
❌	test_chain_function	36.3 µs	41.5 µs	-14.21%
❌	test_chain_two_functions	45.7 µs	51.8 µs	-13.47%
❌	test_chain_nested_functions	45.5 µs	51.8 µs	-13.90%
❌	test_generator_python	43.8 µs	48.9 µs	-11.58%
❌	test_generator_rust	27.3 µs	32.3 µs	-18.16%
❌	test_isinstance_json	26.5 µs	30.4 µs	-14.95%
❌	test_int_error	68.8 µs	80.5 µs	-17.16%
❌	test_definition_in_tree	6 ms	7.3 ms	-21.48%
❌	test_definition_out_of_tree	9.8 ms	11.2 ms	-13.91%
❌	test_model_instance	48 µs	55.7 µs	-16.19%
❌	test_model_instance_abc	48.9 µs	66.7 µs	-36.49%
❌	test_validate_literal[python-few_str_enum]	18.2 µs	22.6 µs	-24.46%
❌	test_validate_literal[python-few_mixed]	19.8 µs	24.3 µs	-22.69%
❌	test_validate_literal[json-few_small_strings]	22.7 µs	26.7 µs	-17.47%
❌	test_validate_literal[json-few_large_strings]	24 µs	27.5 µs	-14.27%
❌	test_validate_literal[json-few_str_enum]	25.6 µs	29.7 µs	-16.09%
❌	test_validate_literal[json-many_small_strings]	22.8 µs	41.6 µs	-82.18%
❌	test_validate_literal[json-many_large_strings]	24.1 µs	28.5 µs	-18.30%
❌	test_validate_literal[json-few_mixed]	21.2 µs	26.3 µs	-24.01%
❌	test_core_root_model	96.3 µs	121 µs	-25.62%
❌	test_int_range	18.8 µs	20.7 µs	-10.30%
❌	test_core_dict	292.7 µs	380 µs	-29.84%
❌	test_core_dict_filter	302.5 µs	390 µs	-28.92%
❌	test_core_json	386.9 µs	453 µs	-17.08%
❌	test_json_direct_list_str	928.8 µs	1,088.4 µs	-17.18%
❌	test_python_json_list_str	716.6 µs	1,007 µs	-40.53%
❌	test_json_any_list_str	1.6 ms	1.8 ms	-10.83%
❌	test_json_direct_list_int	907 µs	1,063.3 µs	-17.23%
❌	test_json_any_list_int	1.6 ms	1.7 ms	-11.12%
❌	test_python_json_list_int	712.6 µs	1,003.2 µs	-40.79%
❌	test_python_json_list_none	692.3 µs	984.9 µs	-42.25%
❌	test_model_list_core_json	975.1 µs	1,131.4 µs	-16.03%

[adriangb]

In my opinion, reliable benchmarks, more consistent performance, and saving developer time chasing down weirdness (see #656 (comment) for context) are well worth a mean 15% performance regression on micro benchmarks. There is also the option of using the default allocator only for benchmarks but keeping MiMalloc for the prod builds which technically might give us the best of both worlds but has the con that benchmark performance would not be reflective of real-world performance (although I think a 10-25% variation between them would be fine as long as it's relatively consistent).

@samuelcolvin what do you think?

[codecov-commenter]

Codecov Report

Merging #662 (ac54212) into main (e7108e8) will not change coverage.
The diff coverage is n/a.

❗ Your organization is not using the GitHub App Integration. As a result you may experience degraded service beginning May 15th. Please install the Github App Integration for your organization. Read more.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #662   +/-   ##
=======================================
  Coverage   93.90%   93.90%           
=======================================
  Files          99       99           
  Lines       13756    13756           
  Branches       25       25           
=======================================
  Hits        12917    12917           
  Misses        833      833           
  Partials        6        6           

Impacted Files	Coverage Δ
src/lib.rs	100.00% <ø> (ø)

---

Continue to review full report in Codecov by Sentry.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update e7108e8...ac54212. Read the comment docs.

[dmontagu]

I thought about it more and realized I think the make build-dev thing doesn't work for me locally if it isn't already pip installed in editable mode. (Which is often an issue for me because I switch between released version and locally installed versions sometimes.) Maybe try:

Suggested change

	make build-prod
	pip install -e .
	make build-prod

[adriangb]

How will we know which version is installed then 🥲

[adriangb]

Similarly, I guess we need to confirm that the MiMalloc version is what wheels are using

[dmontagu]

Not sure if it's worth it, but I would imagine you could add an attribute to the pydantic_core._pydantic_core module conditionally based on the mimalloc feature flag? (And then check for it when uploading the wheel if necessary?)

diff --git a/pydantic_core/_pydantic_core.pyi b/pydantic_core/_pydantic_core.pyi
index 6859a16..517d507 100644
--- a/pydantic_core/_pydantic_core.pyi
+++ b/pydantic_core/_pydantic_core.pyi
@@ -21,6 +21,7 @@ from _typeshed import SupportsAllComparisons
 
 __all__ = (
     '__version__',
+    'allocator',
     'build_profile',
     'SchemaValidator',
     'SchemaSerializer',
@@ -35,6 +36,7 @@ __all__ = (
     'list_all_errors',
 )
 __version__: str
+allocator: str
 build_profile: str
 
 T = TypeVar('T', default=Any, covariant=True)
@@ -167,7 +169,8 @@ class Url(SupportsAllComparisons):
     def query(self) -> 'str | None': ...
     @property
     def fragment(self) -> 'str | None': ...
-    def __init__(self, url: str) -> None: ...
+    @staticmethod
+    def __new__(cls, url: str) -> Url: ...
     def unicode_host(self) -> 'str | None': ...
     def query_params(self) -> 'list[tuple[str, str]]': ...
     def unicode_string(self) -> str: ...
diff --git a/src/lib.rs b/src/lib.rs
index 613220a..9e9229e 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -42,9 +42,20 @@ pub fn get_version() -> String {
     version.replace("-alpha", "a").replace("-beta", "b")
 }
 
+#[cfg(feature = "mimalloc")]
+fn get_allocator() -> String {
+    "mimalloc".to_owned()
+}
+
+#[cfg(not(feature = "mimalloc"))]
+fn get_allocator() -> String {
+    "default".to_owned()
+}
+
 #[pymodule]
 fn _pydantic_core(_py: Python, m: &PyModule) -> PyResult<()> {
     m.add("__version__", get_version())?;
+    m.add("allocator", get_allocator())?;
     m.add("build_profile", env!("PROFILE"))?;
     m.add_class::<PySome>()?;
     m.add_class::<SchemaValidator>()?;

(Haven't tested; also maybe it's overkill)

[adriangb]

Looks like it’s using the default allocator for benchmarks now based just on the slowdowns 😂.

I was thinking something similar but it’s complex. How do you feel about moving forward without that and add ing it later if it’s really needed?

[dmontagu]

@samuelcolvin I'll merge this now since @adriangb and I agree it's good and it sounded earlier like you were okay with this, but just comment if you want more work done or whatever

Edit: oh looks like @adriangb beat me