Pydantic benchmarks

Benchmark performance and explore the new features of various releases of Pydantic.

The benchmarks and their implications are explored in more detail in the following blog posts:

• The first blog post showed how simply changing a few lines of code from the v1 API to v2, yielded a 5x speedup, thanks to the underlying pydantic-core being rewritten in Rust 🦀 💪🏽

• The second blog post shows that, with the right knowledge of Pydantic v2's new features, it's possible to get a 10x or greater performance improvement over v1. Additionally, each successive major release of Pydantic v2 has shown incremental improvements in performance, thanks to underlying optimizations and innovations at the Rust level. More such improvements are likely in future versions!

See the tables below showing how the performance of Pydantic v2 has steadily improved with each release.

Setup

Activate a Python virtual environment and install the dependencies as follows.

# Assuming that the uv package manager is installed
# https://github.com/astral-sh/uv
uv venv
source .venv/bin/activate
uv pip install pydantic

Install any specific versions of Pydantic as required by specifying their major and minor version numbers to benchmark each of them.

# uv pip install pydantic==2.0.3
# uv pip install pydantic==2.1.1
# uv pip install pydantic==2.2.1
# uv pip install pydantic==2.3.0
# uv pip install pydantic==2.4.2
# uv pip install pydantic==2.5.2
# uv pip install pydantic==2.6.0
# uv pip install pydantic==2.7.0
# uv pip install pydantic==2.8.2
uv pip install pydantic==2.9.0

Navigate to the respective directories v1 and v2 to run the benchmark for each version of Pydantic.

Dataset

The dataset used for this benchmark is the Wine Reviews dataset from Kaggle, containing ~130k reviews on wines along with other metadata. For convenience and ease of reproduction, the dataset is made available here as line-delimited JSON in the data directory.

Benchmark results

The benchmark was run using pytest-benchmark. See the respective version directories for steps to reproduce the benchmark results.

Note

• The timing numbers below are from a 2023 M3 Macbook Pro with 32GB RAM
• The times reported are an average over 10 runs, with 5 warmup runs

Basic validator

The basic validator's schema makes use of simple Pydantic models, with very minor changes between v1 and v2.

Version	Run time (sec)	Speedup factor over v1
1.10.14	2.829	1.0
2.0.3	0.567	5.0
2.1.1	0.595	4.8
2.2.1	0.563	5.0
2.3.0	0.556	5.1
2.4.2	0.545	5.2
2.5.3	0.626	4.5
2.6.0	0.492	5.7
2.7.0	0.475	6.0
2.8.2	0.418	6.8
2.9.0	0.415	6.8
2.10.2	0.424	6.7

Improved validator

The improved validator's schema makes use of intermediate concepts in Pydantic, using some of the new features available in v2. It produces results that are identical to the basic validator. Note that in this case, we cannot test v1 because the features used are available only in v2.

Version	Run time (sec)	Speedup factor over v1
2.0.3	0.249	9.1
2.1.1	0.253	8.9
2.2.1	0.253	8.9
2.3.0	0.250	9.0
2.4.2	0.236	9.6
2.5.3	0.273	8.2
2.6.0	0.205	10.9
2.7.0	0.193	11.6
2.8.2	0.175	12.8
2.9.0	0.162	13.9
2.10.2	0.155	14.3

Tip

With the right knowledge of Pydantic v2 features, it's possible to optimize and tune your validation workflow to a much greater extent. The performance gains are largely due to underlying improvements to pydantic-core and PyO3 at the Rust level, that are then noticeable at Python level.