Speed up message sorting in write_mo() and sorted write_po() - #1438

write_mo() always sorts the catalog's messages, and write_po(sort_output=True) does too. list.sort() calls Message.__lt__, which goes through __cmp__ and rebuilds both (id, context) tuples (plus a nested helper function) for every comparison, so a catalog of n messages builds about 2·n·log2(n) tuples. On real catalogs the sort is a large part of pybabel compile.

This adds a small private helper, _sort_messages_by_id(), used by write_mo() and pofile._sort_messages(). For plain Message objects whose id (or first plural id) and context are str/None, it builds the same (id, context or '') key that __cmp__ compares, once per message, and does a key sort. Python's sort is stable either way, so ties keep their order and the output is the same. If any message is a Message subclass or has an unusual id/context, it falls back to messages.sort(), so custom comparison methods and the existing errors are unchanged.

Benchmarks

Real catalogs: Sphinx (ar, ca, fi, hu, sv, zh_TW; 869 messages each) and Django 5.2.8 (de, ja; 348 messages each), all eight per sample. Each sample was a separate process pinned to one core on Linux. master and this branch ran in alternating order (AB, BA, ...); the table gives the median per-pair ratio over 8 pairs on 3.12 and 6 on 3.15 and 3.9.

write_mo alone (catalogs already parsed) goes from 16.0 ms to 5.0 ms on 3.12 (3.2x, 6 pairs). The default write_po() (no sorting) doesn't change.

Checks

• I compared master and this branch in separate processes: write_mo (with and without use_fuzzy) and write_po (sorted by message, by file and unsorted) for the 8 real catalogs and 6 generated catalogs with plurals, contexts (None, '', mixed case, non-ASCII) and duplicate ids, plus subclass and invalid-id cases (empty plural tuple, None id, non-string context). Output bytes, and exception types and messages, were identical on Python 3.9, 3.12 and 3.15.
• New tests in tests/messages/test_catalog.py check the order against sorted() (plurals, contexts, ties) and that a subclass's __cmp__ is still used.
• The full test suite passes on Python 3.9, 3.12 and 3.15, and pre-commit run --all-files passes.

Related: #1374 (open) adds a second sort of the MO offset table by encoded key further down in write_mo(). It doesn't touch the message sort, so the two changes are independent and don't conflict.

🤖 Generated with Claude Code