[arieleven]

Requested by Taltal · project thread

Problem

The filesystem transport names each message file <ms>_<uuid4>.<queue>.msg, where <ms> is int(round(monotonic() * 1000)), and _get() takes the first name after sorted(os.listdir(...)). That is a text sort, so:

• messages published within the same millisecond come back in uuid order, which is random. A producer that publishes 20 messages in a loop gets them back as, for example, [0, 5, 4, 2, 3, 1, 6, 12, 7, 10, ...];
• the order also breaks when the millisecond counter gains a digit, because 10000_... sorts before 9999_....

Fix

Zero-pad the millisecond prefix and add a per-process sequence number (itertools.count()) between it and the uuid. Within one process the names now sort in publish order. Across processes, order is still by millisecond, as before. _get(), _size() and _purge() are unchanged: they still match on what follows the first ..

Upgrade note: while files with the old unpadded names are still in the folder, new files (with a leading 0) sort before them. Old messages are still delivered; only their order relative to new ones changes during the switch.

Tests

test_messages_are_consumed_in_publish_order patches monotonic for two cases: ten messages in the same millisecond, and the counter going from 9999 to 10004. Both fail on main (7dca15a6) and pass with the fix.

• t/unit/transport/test_filesystem.py: 221 passed.
• t/unit (with --continue-on-collection-errors, because the SQS/Azure extras are not installed here): 1573 passed, 186 skipped. One failure, test_FilesystemOpenErrors::test_get_closes_file_when_read_fails, fails on main too, and only in the full run (it passes when the file runs alone).
• flake8 and pydocstyle are clean on the changed files. mypy kombu/transport/filesystem.py reports the same 22 errors as on main.

I used Claude Code to help find and write this fix. I reviewed and tested the change.