[kinyoklion]

Requirements

• I have added test coverage for new or changed functionality
• I have followed the repository's pull request submission guidelines
• I have validated my changes against all supported platform versions
• Sync/async parity: matching changes made to async_* siblings (or N/A). N/A: the file watcher serves both clients.

Related issues

This resolves the SDK findings of the Windows stability assessment for the overrides work: plans/python-overrides-windows-test-stability.md in sdk-scratchpad.

• Finding 1, the flaky retry test, is fixed by test: Change the file size in the override retry test so the poller sees the malformed write #576.
• Findings 2, 3 and 4 are fixed here.

Describe the solution you've provided

All three changes are in Watcher, in ldclient/impl/integrations/files/filedata.py, which the file-based override source uses in watching mode.

1. watchdog 7 delays Windows events by a second, and starves them in a busy directory (Finding 2).
   • watchdog 7.0.0 (2026-10-31) changed its Windows reader. It now waits for the observer timeout after the first notification before delivering anything, and it delivers nothing while notifications keep arriving within that timeout.
   • With the default one-second timeout, a change to an override file was reported about a second late on Windows. It was not reported at all while another file in the same directory, such as a log, changed more often than once a second.
   • The assessment measured, on Windows Server 2025 with watchdog 7.1.0: a median latency of 1003 ms, and three of three changes not reported within 8 s next to a file appended every 200 ms. With a 100 ms timeout: 102 to 110 ms, and 3 of 3 reported.
   • The watcher now builds the observer with timeout=0.1. On Linux, the timeout only sets how often the idle inotify loop wakes.
   • Python 3.11 and later install watchdog 7 under watchdog>=3.0.0, so this affects any Windows deployment that selects ChangeDetection.WATCHING.
2. A touch no longer reloads with watchdog 7 (Finding 3).
   • watchdog 7 reports a change to a file's metadata, such as a new modification time from os.utime, as an attrib event. Earlier versions reported it as modified.
   • The watcher ignored attrib, so touching a file to force a reload stopped working in watching mode, while it still works in polling mode.
   • attrib now counts as a change. A plain read does not produce it, so the reloads stay quiet. The reloader's content digest still skips applying an unchanged file.
3. watchdog 6 on Windows can miss a change right after a watch starts (Finding 4).
   • watchdog 6.0.0, the last version that supports Python 3.10, returns from starting a watch after a fixed 10 ms. Its reader thread may not yet have asked for the first notification, and anything before that request is lost.
   • The assessment found that a deletion made immediately after the watcher was created was reported 0 of 10 times, and 10 of 10 times after one thread yield.
   • The watcher now yields once after each watch it sets up.

Tests in test_filedata.py:

• a change is reported within 0.5 s;
• a change is reported within 2 s while another file in the directory is appended to five times a second;
• a change of the modification time alone is reported.

Describe alternatives you've considered

• Pinning watchdog<7 in the test-filesource extra. Applications bring their own watchdog, so a pin would only hide the behavior from CI.
• Running a slow safety poll under watching mode, so that a lost or starved notification is bounded. That is a design decision for the spec, not part of this fix.
• Documenting that watching mode ignores a touch, instead of reacting to attrib. That would leave watching and polling behaving differently for the same file.

Additional context

Testing:

• On Linux with watchdog 7.1.0, the touch test fails without this change and passes with it.
• The latency and busy-directory tests pass on Linux either way, because inotify delivers each event at once. Windows CI runs them against watchdog 7, which they target.
• Every watcher test passes with watchdog 6.0.0 and with watchdog 7.1.0. That includes the existing test that a read alone does not signal.
• The file data and override source modules pass on watchdog 7.1.0 (263 tests).
• The full suite passes (2,120 tests), without the Redis, DynamoDB and Consul store tests, which need servers that were not running.
• mypy, isort and pycodestyle are clean.

This PR merges cleanly with #574 and #576.

Not in this PR, from Finding 6 of the assessment. None of these come from the overrides work:

• a latency assertion in test_async_migrator.py that is too tight for Windows clocks;
• a streaming retry test that is flaky on main;
• the big segments status-polling contract test, which turns Linux CI on feat/overrides red in 4 of 5 runs and needs its own investigation before the merge into main.

The assessment also recommends reporting the watchdog 7 behavior upstream; that is not done yet.

_{Opened from termag}

---

Note

Overview
Updates the file Watcher in watching mode so override/data reloads behave reliably with watchdog 6–7, especially on Windows.

The observer now uses a 100 ms delivery timeout instead of watchdog’s 1 s default, so changes are reported quickly and are not indefinitely delayed when another file in the same directory keeps notifying. attrib is treated as a content-relevant event so a mtime-only touch triggers a reload again under watchdog 7. After each directory watch is scheduled, a thread yield (time.sleep(0)) avoids missing an immediate post-start change on watchdog 6 for Windows.

New watcher tests cover sub-second latency, detection alongside a busy log file, and mtime-only updates.

^{Reviewed by Cursor Bugbot for commit 546a7f5. Bugbot is set up for automated code reviews on this repo. Configure here.}