Speed up model field reads ~4.6x (FieldDescriptor)

[wbarnha]

Description

FieldDescriptor defines both __get__ and __set__, so it is a data descriptor: every read of every field of every model instance goes through it. This is two changes, and the pure-Python one turns out to matter more than the Cython.

1. Two cached_property declarations on the read path (pure Python)

lazy_coercion and related_models were mode.utils.objects.cached_property. That decorator defines __set__, which makes it a data descriptor — so it intercepts every read even once the value is cached. lazy_coercion is read on every field access:

cached_property read : 166.9 ns
plain attribute read :  32.9 ns

Both are now ordinary attributes, assigned by on_model_attached() where _expr becomes available. related_models is only read at class-definition time; neither appears in as_dict(), so clone() is unaffected.

Worth flagging for review: this had to be fixed in two places. FieldDescriptor declared them, but so did FieldDescriptorT in faust/types/models.py. Removing only the subclass's copy changed nothing measurable, because the base class's data descriptor still intercepted the read — an easy thing to half-fix.

2. faust/models/_cython/fields.pyx — __get__ on an extension type

Follows the pattern already in the tree: pure-Python implementation stays and is used whenever the extension could not be built or NO_CYTHON is set.

Deliberately minimal. Only __get__ moves, and only the four attributes it touches are declared on the extension type. Everything else remains an ordinary Python attribute in __dict__, so as_dict()/clone(), tests that set attributes directly, FieldDescriptor[str] subscripting and isinstance(x, FieldDescriptorT) are all unaffected.

Benchmarks

Medians of 5 interleaved runs on CPython 3.11, with a raw __dict__ read as a control for machine noise:

metric	master	pure-py only	+cython	py vs master	cython vs master
1 field read	321 ns	171 ns	70 ns	1.88x	4.59x
5 field reads	1620 ns	857 ns	287 ns	1.89x	5.65x
from_data	4227 ns	3865 ns	4287 ns	1.09x	0.99x
raw __dict__ (control)	60 ns	64 ns	64 ns	—	—

Reads land essentially on the raw-dict floor.

The honest caveat is construction. The pure-Python change alone makes from_data 1.09x faster; the extension gives that back, because cdef public compiles to getset descriptors and the still-Python write path (__set__, should_coerce) reads those attributes through them — measurably slower than a plain __dict__ hit (45.7 ns vs 33.4 ns for descriptor.field).

I tried moving __set__ into the extension as well so both hot paths would use C-level attribute access. It measured worse, so it was reverted rather than kept on the theory that it should have helped.

Net: for a model whose fields are read at least once, this is comfortably ahead (5847 ns → 4574 ns for construct-plus-read-all); for one constructed and never read, it is a wash. If reviewers would rather have the construction win than the read win, dropping the .pyx and keeping change 1 is a coherent option.

Testing

• New tests exercise the read path against both implementations: class vs instance access, lazy coercion of a nested model (including that the coerced value is cached), optional fields, arbitrary attribute assignment and clone().
• A regression test asserts neither attribute resolves to a cached_property anywhere in the MRO — that is the actual invariant, rather than where the value happens to be stored (the extension type keeps lazy_coercion as a C struct member, the Python one as an instance attribute).
• Full suite passes both ways: 2220 with the extension, 2214 with NO_CYTHON=1.
• flake8 / black / isort clean; sdist ships the new .pyx.

Relationship to other PRs

Independent of #751 (offset/scheduler/sensor accelerators) — different files, no overlap. This one was identified there as the remaining promising candidate and deferred as too large for that PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_019oX4oGCQGgvPJHabjwfaBk

---

Generated by Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.04%. Comparing base (803c7a4) to head (562b623).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #753      +/-   ##
==========================================
- Coverage   96.06%   96.04%   -0.02%     
==========================================
  Files         103      103              
  Lines       11072    11076       +4     
  Branches     1191     1191              
==========================================
+ Hits        10636    10638       +2     
- Misses        345      347       +2     
  Partials       91       91              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.