Performance & Benchmarks¶

PolyXML is engineered to process gigabytes of XML per second by leveraging Rust's zero-cost abstractions, quick-xml streaming events, and lexical-core numeric conversions.

The repository includes a fully reusable, automated benchmark suite covering both pure Rust Criterion tests and Python comparative benchmarks.

1. Python Deserialization & Serialization Throughput¶

Benchmarks conducted using Python 3.12 (abi3-py312) across 10,000-element streaming payloads (~724 KB XML), micro sensor payloads (~100B), and enterprise orders:

Batch Catalog Workload (10,000 items, ~724 KB XML)¶

Key Takeaway: - Vs Typed Dataclasses (xsdata): PolyXML is 16.0x faster at deserialization and 38.7x faster at serialization, while saving over 1.2 MB of memory. - Vs Dict Parsers (xmltodict): PolyXML is 4.1x faster on deserialization and 10.8x faster on serialization, while instantiating strongly-typed dataclasses instead of unstructured string dictionaries. - Vs C Proxies (lxml.objectify): PolyXML executes within 1.39x of raw C dynamic proxy trees, while returning genuine Python dataclasses with IDE autocomplete and type safety.

Real-Time Micro Telemetry Workload (Sensor ~100B, UCI Telemetry)¶

Critical telemetry commands and sensor packets deserialize in 2.5 microseconds, beating even C-based DOM parsers (lxml at 3.1 μs, lxml.objectify at 3.3 μs).

2. Pure Rust Core Throughput (crates/polyxml-core)¶

Statistical benchmarks measured with Criterion.rs:

3. Key-Value Storage & Binary IPC Throughput (polyxml.dumps_binary / loads_binary)¶

When storing XML dataclasses and Pydantic models in embedded transactional key-value databases (libmdbx, LMDB, RocksDB) or communicating over Unix Domain Sockets/multiprocessing queues, Python's traditional cloudpickle and standard pickle encounter severe GIL and GC bottlenecks.

polyxml.dumps_binary() and polyxml.loads_binary() provide high-speed MessagePack encoding with universal leaf type hooks (XmlDate, XmlDateTime, XmlDuration, XmlTime, Decimal, QName, Enum, Path, and Pydantic models):

10,000 Complex XML Entities in Real MDBX Pipeline¶

Key Takeaways:¶

• 7.9x Faster Serialization: Slashes per-object serialization from 48.5 μs down to 6.1 μs.
• 53.9% Storage Space Reduction: Cuts stored byte size from 547 bytes to 252 bytes per entity.
• 3.2x MDBX Transaction Speedup: Real database writes into libmdbx jump from 18,287 ops/s to 58,781 ops/s.
• 100% Fidelity Guarantee: Round-trips preserve exact dataclass types, field metadata, and XML primitive representations with assert loads(dumps(x)) == x.

4. How to Reproduce Benchmarks¶

The benchmark suite is reusable and version-controlled.

Run All Benchmarks (Rust + Python)¶

./benchmarks/run_all.sh

Run Pure Rust Criterion Benchmarks¶

cargo bench --bench core_benchmarks

Run Python Comparative Benchmarks CLI¶

python -m benchmarks --workload all --catalog-sizes 1000 10000 --iterations 25 --output-md benchmarks/results.md --output-json benchmarks/results.json