[John-Cusack]

Pull Request Checklist

• 🧪 Added tests for changed code.
• 🛠️ All tests pass when run locally (run tox or make test to find out!).
• 🧹 All linting checks pass when run locally (run tox -e lint or make lint to find out!).
• 📚 Updated documentation for changed code.
• ⏩ Code is up-to-date with the master branch.

No documentation update: there is no public API change — arrow.get still
builds the same parsers with the same results; the memo is internal.

Description of Changes

Every arrow.get(s) / arrow.get(s, fmt) constructs a fresh
DateTimeParser(locale) with cache_size=0, so each parse pays:

1. DateTimeParser.__init__ — locale lookup + 6 _generate_choice_re →
   re.compile calls, and
2. _generate_pattern_re(fmt) — a full regex rebuild from the format
   string (the top cProfile frame of a format parse).

The existing lru_cache opt-in (cache_size > 0, default 0) cannot help
this path: it wraps the bound method per instance, and the factory
discards the instance after one call.

This PR:

• memoizes _generate_pattern_re at class level in arrow/parser.py
  (bounded at 128 entries, GIL-atomic dict access under a lock for
  eviction). The key embeds a fingerprint of the parser's effective
  input map (parser class + locale name + token→pattern/flags snapshot
  built in __init__), so per-instance or per-subclass map
  customizations — including the test_parse_unrecognized_token
_BASE_INPUT_RE_MAP mock — can never read another parser's entry.
  Compiled patterns are immutable and (map, fmt)-determined, so the
  cache is semantically transparent: identical re objects, identical
  matches and errors.
• reuses one DateTimeParser per locale string in the factory
  (arrow/factory.py _get_parser), so __init__'s locale-regex
  rebuild also runs once per locale instead of once per call. Parsers
  are effectively immutable after construction for stock locales.

Measured

Ryzen 9 5950X, CPython 3.12.3, frozen 2000-item batches
(YYYY-MM-DD HH:mm:ss format batch + ISO-8601 batch), each arm a fresh
subprocess, in-process min-of-5 after warmup, min across 3 processes;
peak RSS via /proc VmHWM (this box's ru_maxrss is constant-broken —
verified against VmHWM, identical values both arms):

Track	Stock (master @ 2224255)	Patched	Speedup
arrow.get(s, FMT) × 2000	0.0441 s	0.0114 s	3.86x
arrow.get(s) ISO × 2000	0.0652 s	0.0228 s	2.86x
representative pipeline (parse→shift→format→humanize) × 2000	0.0709 s	0.0391 s	1.81x

Peak RSS flat across arms (~19.0 MB both). Outcome counts identical
both arms (all 2000 items parse in every track on both sides).

Reproduce (at 2224255 vs this branch, venv per requirements/requirements-tests.txt):

pip install -r requirements/requirements-tests.txt && pip install -e .
python bench_arrow.py   # frozen 2000-item fmt/iso/pipeline tracks above

(Bench script available on request — say the word and I'll attach it.)

Test / lint proof

Their own gates, venv built per Makefile
(requirements-tests.txt + pip install -e .):

• pytest: 1908 passed (4 new: cross-instance memo sharing,
  locale isolation en/fr/de, 128-entry eviction bound + transparent
  rebuild, factory one-parser-per-locale). Pre-existing caching tests
  (test_parser_no_caching, test_parser_1_line_caching, ...) and the
  _BASE_INPUT_RE_MAP mock test pass unmodified.
• Coverage gate (--cov-branch --cov=arrow --cov-fail-under=99, as in
  tox.ini): 99.93%, required 99% reached.
• pre-commit run --all-files: all hooks pass (black, isort,
  flake8 + bugbear + annotations, mypy, pyupgrade).
• Differential equivalence: 20-case corpus (ISO, 14 format shapes incl.
  locale month/day names, week dates, meridiem, fractional seconds, plus
  5 error-path cases) through arrow.get — stock vs patched
  20/20 identical, including exception types and messages.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.00%. Comparing base (2224255) to head (af2999f).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff            @@
##            master     #1398   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files           10        10           
  Lines         2315      2342   +27     
  Branches       358       360    +2     
=========================================
+ Hits          2315      2342   +27     

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

[John-Cusack]

CI triage note (for reviewers): 24/27 legs green, including Linting and all CPython 3.10–3.14 legs. The 3 reds are not from this change:

• windows-latest (pypy-3.11) fails its tox run identically on unmodified master (2224255 — visible across several master check-suite runs), so pre-existing.
• ubuntu-latest (3.8) / ubuntu-latest (3.9) fail in the "Set up Python" step before any test executes (EOL-Python provisioning), likewise unrelated to this diff.

Happy to rebase/retest if you want a fresh full-green signal.

[John-Cusack]

Closing this. The shared pattern cache is keyed on a snapshot of the input map taken in __init__. If a parser's _input_re_map is changed after that, it and a stock arrow.get("2013", "YYYY") can break each other: whichever parses YYYY first fills the cache entry, and the other then fails. The gain for a real caller is also small: in ics 0.7.3, parsing a generated 500-event calendar was less than 5% faster.