benchmark

DAG-CBOR encode/decode benchmark across Python implementations.

Compared:

• libipld (Rust)
• cbrrr (C)
• py-ipld-dag (Python wrapper over cbor2 with canonical=True; cbor2 itself is Rust)
• dag_cbor (pure Python, used as the 1× baseline)

Fixtures: canada.json, citm_catalog.json, github.json, twitter.json (loaded from ../data/, parsed once, then encoded to DAG-CBOR via libipld for the decode benchmarks).

Run

./run.sh           # encode + decode
./run.sh encode    # encode only
./run.sh decode    # decode only

Outputs serialization.png and deserialization.png next to chart.py. Raw history goes into .benchmarks/ (pytest-benchmark autosave).

Which libipld is measured?

By default, requirements.txt pulls the published PGO-optimized wheel from PyPI. This is what users actually get when they pip install libipld, so the charts reflect real-world performance.

To benchmark a locally-built PGO wheel instead, edit the libipld line in requirements.txt:

libipld @ file:///abs/path/to/libipld-*.whl

Building a local PGO wheel is out of scope here.

Filter

Skip a slow library (e.g. pure-Python dag_cbor on canada):

uv run --with-requirements requirements.txt --with-editable .. \
    pytest --benchmark-enable --benchmark-json=results.json -k "not dag_cbor"
uv run --with-requirements requirements.txt python chart.py results.json