Skip to content

python-json-binary-aot-2026-09.md

Latest commit

 

History

History
122 lines (91 loc) · 8.21 KB

python-json-binary-aot-2026-09.md

File metadata and controls

122 lines (91 loc) · 8.21 KB
Raw
title Python JSON, Binary, and AOT Benchmarks (September 2026)
description Verified, reproducible performance benchmarks for Python JSON data-binding, binary serialization, and Ahead-of-Time (AOT) PyO3 C-extensions with committed raw outputs.

Python JSON, Binary, and AOT Benchmarks — September 2026

This report provides reproducible performance evidence for Python JSON data-binding, binary serialization, and Ahead-of-Time (AOT) compiled native extensions. Every measurement is backed by committed harnesses, identical semantic model checks, and raw sample logs.

Environment and Methodology

Item Value
Run date 2026-09-26 (America/Chicago)
Host AMD Ryzen 5 4500, 12 logical CPUs, 7.7 GiB RAM
OS Ubuntu 26.04.1 under Linux 6.18.33.2 WSL2, x86-64
Toolchains Python 3.12.14, PolyXML 0.24.0 (PyO3 abi3-py312), Rust 1.98.1
Competitors & Libraries xsdata 26.2, cloudpickle 3.1.2, lz4 4.4.5, standard library pickle (protocol 5)
Raw Data Directory docs/benchmarks/data/2026-09-26/

1. JSON Data-Binding: PolyXML vs xsdata

This suite compares PolyXML native JSON APIs (polyxml.loads_json, dumps_json, JsonParser, JsonSerializer) against xsdata JSON data-binding on the exact same Python dataclass models and identical JSON payloads.

Deserialization (JSON → Typed Dataclass)

Workload Payload Size xsdata Median PolyXML Native Median Speedup vs xsdata PolyXML Throughput
Sensor (Small) 128 B 156.5 μs 12.8 μs 12.2x faster 9.5 MB/s
Order (Nested) 793 B 1,284.5 μs 140.2 μs 9.2x faster 5.4 MB/s
Batch (100 Items) 12,842 B 20,570.3 μs 1,326.7 μs 15.5x faster 9.2 MB/s

Serialization (Typed Dataclass → JSON)

Workload Payload Size xsdata Median PolyXML Native Median Speedup vs xsdata PolyXML Throughput
Sensor (Small) 128 B 86.4 μs 10.3 μs 8.4x faster 11.9 MB/s
Order (Nested) 793 B 505.6 μs 60.4 μs 8.4x faster 12.5 MB/s
Batch (100 Items) 12,842 B 6,586.1 μs 711.7 μs 9.3x faster 17.2 MB/s

2. Binary Serialization: PolyXML vs pickle and cloudpickle

This suite compares PolyXML Binary (polyxml.dumps_binary, loads_binary based on MessagePack) with standard library pickle (protocol 5 with C accelerator) and cloudpickle across direct uncompressed and LZ4 compressed modes.

Direct (Uncompressed) Serialization

Workload Serializer Payload Size Encode Median Decode Median Encode Ops/s Decode Ops/s
Sensor (Small) PolyXML Binary 118 B 5.3 μs 3.6 μs 188,147 276,129
Pickle 5 (Stdlib C) 185 B 5.1 μs 8.2 μs 195,695 121,721
CloudPickle 185 B 18.8 μs 8.1 μs 53,233 123,678
Order (Nested) PolyXML Binary 642 B 23.6 μs 16.7 μs 42,382 59,893
Pickle 5 (Stdlib C) 684 B 24.2 μs 27.3 μs 41,328 36,668
CloudPickle 684 B 92.7 μs 27.3 μs 10,792 36,621
Batch (100 Items) PolyXML Binary 11,853 B 180.4 μs 331.0 μs 5,544 3,022
Pickle 5 (Stdlib C) 7,192 B 248.7 μs 275.2 μs 4,021 3,634
CloudPickle 7,192 B 718.1 μs 282.2 μs 1,392 3,544

Compressed (LZ4) Serialization

Workload Serializer LZ4 Size Encode+LZ4 Median Decode+LZ4 Median Encode+LZ4 Ops/s Decode+LZ4 Ops/s
Sensor (Small) PolyXML Binary 138 B 6.9 μs 4.2 μs 144,959 239,349
Pickle 5 (Stdlib C) 198 B 6.9 μs 9.7 μs 144,020 103,157
CloudPickle 198 B 20.5 μs 9.9 μs 48,807 100,873
Order (Nested) PolyXML Binary 344 B 26.9 μs 17.4 μs 37,236 57,511
Pickle 5 (Stdlib C) 443 B 27.6 μs 30.3 μs 36,294 33,055
CloudPickle 443 B 90.2 μs 31.0 μs 11,086 32,218
Batch (100 Items) PolyXML Binary 1,937 B 203.5 μs 341.9 μs 4,913 2,925
Pickle 5 (Stdlib C) 1,986 B 265.2 μs 508.1 μs 3,771 1,968
CloudPickle 1,986 B 721.9 μs 301.8 μs 1,358 3,314

3. Real-World Ahead-of-Time (AOT) PyO3 C-Extension vs Standard Dataclass

This suite evaluates the performance difference between PolyXML's standard Python @dataclass runtime (polyxml.deserialize, serialize) and Ahead-of-Time compiled PyO3 native extensions (--backend aot).

  • Workload 1: Real-world USAF UCI v2.5 command and control entity telemetry (benchmarks/python-aot/data/uci_entity.xml, 1,510 bytes XML). Models: uci_entity_core.EntityMt dataclass vs uci_aot.EntityMt PyO3 extension.
  • Workload 2: Synthetic SensorReading telemetry (226 bytes XML).
  • Runner: benchmarks/python-aot/
  • Raw output: Text log · JSON metrics
  • Mathematical Consistency: Verified $\text{ops/sec} \times \text{median latency (s)} = 1.0$ exactly across all measurements.

Deserialization (XML → Model)

Workload XML Size Dataclass Latency Dataclass Ops/s AOT Latency AOT Ops/s AOT Speedup AOT Throughput
USAF UCI v2.5 Telemetry (Real-World) 1,510 B 62.5 μs 16,003 7.44 μs 134,336 8.39x 193.5 MB/s
SensorReading (Synthetic) 226 B 11.3 μs 88,367 2.31 μs 433,839 4.91x 93.5 MB/s

Serialization (Model → XML)

Workload XML Size Dataclass Latency Dataclass Ops/s AOT Latency AOT Ops/s AOT Speedup AOT Throughput
USAF UCI v2.5 Telemetry (Real-World) 1,510 B 39.8 μs 25,141 3.88 μs 257,932 10.26x 371.4 MB/s
SensorReading (Synthetic) 226 B 7.4 μs 135,980 1.74 μs 573,393 4.22x 123.6 MB/s

4. Key Takeaways & Tradeoffs

  1. JSON Data-Binding:
    • PolyXML native JSON deserialization is 9.2x to 15.5x faster than xsdata, and serialization is 8.4x to 9.3x faster, while operating on the exact same dataclass models.
  2. Binary Serialization:
    • For small and nested objects, PolyXML MessagePack decoding is 2.3x faster than pickle (3.6 μs vs 8.2 μs for Sensor; 16.7 μs vs 27.3 μs for Order) and creates smaller payloads (118 B vs 185 B).
    • For large homogeneous 100-item batches, pickle's specialized C array packing achieves a smaller uncompressed payload (7.2 KB vs 11.8 KB) and slightly faster decoding (275.2 μs vs 331.0 μs). With LZ4 compression, PolyXML achieves smaller payloads (1,937 B vs 1,986 B).
  3. Ahead-of-Time (AOT) C-Extensions:
    • On production USAF UCI XML telemetry messages, AOT PyO3 native extensions parse XML in 7.44 μs (134,336 ops/sec, 193.5 MB/s), an 8.39x speedup over runtime dataclass reflection.
    • Per-packet latency is measured at 7.44 μs (not sub-microsecond). While extraordinarily fast, we strictly avoid rounding or exaggeration.