Allocations overhead ?

[mselee]

First of all, thanks for the fantastic library.

While benchmarking pydantic v2 on some complex models with a lot of fields/nesting (unfortunately I can't share the models) I noticed abnormal performance due to allocations/deallocations:

@model_validator(mode='wrap') is used in these models but is dwarfed by the Clone and Drop calls. The (de-)allocations represent ~75% of the total run time.

I tried replacing the Box<CombinedValidator> with an Arc<RwLock<CombinedValidator>> for:

• FunctionWrapValidator.validator
• FunctionBeforeValidator.validator
• FunctionAfterValidator.validator

After rebuilding with the patch, the (de-)allocations calls dropped:

Benchmark (1000 iterations)

Container	Time	Allocations	Memory (Peak)
Box<CombinedValidator>	~13 sec	202000	66.4 KiB
Arc<RwLock<CombinedValidator>>	~3 sec	2000	1.6 KiB

But I'm not sure if this is actually sound (ref):

I would suggest using Arc where you are using Box, but unfortunately Validator::set_ref takes &mut self. This prevents calling it on Arc. While we could use Arc<RwLock> and acquire locks as needed, it would still likely be incorrect since currently all cloned validators are entirely independent, while with the Arc approach modifying the validator will also modifies everything which references it.

There's no set_ref in the code base anymore, but I guess the same concern is still valid for validator.complete() ?

Selected Assignee: @samuelcolvin

[samuelcolvin]

Thanks so much for creating the issue, really interesting.

We were discussing Box just the other day and assuming it's overhead was tiny.

@davidhewitt what do you think?

[samuelcolvin]

What was your benchmark of? Validation or validator construction?

[mselee]

It is something similar to this

class ComplexPydanticModel(BaseModel):
    # I also tried `defer_build=True` or removing some of `validate_xxx` but didn't make a noticeable difference
    model_config = ConfigDict(from_attributes=True, validate_default=True, validate_return=True)

    # fields/validators are omitted

class Response(BaseModel, Generic[T]):
    data: T


BenchmarkModel = Response[List[ComplexPydanticModel]]
data = {"data": data}

for _ in range(1000):
    BenchmarkModel.model_validate(data)

data is a list of complex ORM objects and from my testing the list size doesn't matter. What matters is the iterations in the loop i.e. for 100 iterations the runtime is barely affected, but it starts to be noticeable if you increase it:

Iterations	Time
1000	~13 sec
500	~6 sec
100	~3 sec

[davidhewitt]

Thanks for the report and the proposed patch. Inspired by it I'm going a bit further and reducing away many cases where we clone validators (using probably just Rc<CombinedValidator> as we don't need the mutability any more).

I've got a patch which is almost there but I'm aiming to merge #867 first to make the patch simpler.