Skip to content

Performance

Methodology

Benchmarks are reproducible via benchmarks/bench.py (spec 11). Each operation is timed with timeit over N=100_000 iterations, repeated 5 times; the mean and median microseconds-per-operation are reported. The environment (Python version, platform, CPU) is captured in the results JSON (benchmarks/results/results.json).

Three implementations are compared:

  1. plain dict
  2. pure-Python AttributeDict (attributedict._reference — the spec oracle)
  3. C AttributeDict (attributedict)

Measured results (CPython 3.13.13, x86_64; 2026-08-15)

operation dict (µs) reference (µs) C (µs)
construct kwargs 3.0 33.5 3.4
construct mapping 0.5 30.0 3.1
construct nested — 8.8 2.6
getitem 0.074 0.09 0.066
getattr — 0.234 0.071
setitem 0.105 0.119 0.117
setattr — 0.612 0.114
iter 0.482 0.678 0.693
copy (shallow) 1.76 31.9 1.89

Interpretation

Where the C extension improves things (vs the pure-Python reference):

  • Attribute get/set: ~3–5× faster (C avoids the Python __getattribute__ dispatch and str.isidentifier call per lookup).
  • Construction: ~10× faster (recursive conversion is done in C, not Python).
  • copy: ~17× faster.
  • dict.items(d) / mapping views unaffected.

Where it is on par with plain dict:

  • d[key], d[key] = v, del d[key], iteration: essentially identical to plain dict (the operations inherit directly from the dict base).
  • Construction from an existing mapping is slightly slower than plain dict (3.1 µs vs 0.5 µs) because the C code performs the recursive conversion pass at construction (O(n)). This is the documented cost of the conversion feature.

Where it does NOT help:

  • Iteration and mapping views: same as dict (inherited).
  • No claim of "2× faster than dict" is made: for inherited operations the C type is at parity with dict, and the speedups are against the pure-Python baseline, which is the honest comparison.

Construction cost

Construction is O(n) in the number of contained items due to recursive conversion; nested/cyclic structures are handled cycle-safe. This is documented in nested.md and confirmed by the construction benchmarks above.

Reproducing

python benchmarks/bench.py # table
python benchmarks/bench.py --json out.json # JSON data

See benchmarks.md for details.