(beam) branch master updated: Clarify known restrictions for model updates. (#40600)

[tv...@apache.org]

This is an automated email from the ASF dual-hosted git repository.

tvalentyn pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/beam.git


The following commit(s) were added to refs/heads/master by this push:
     new a17e73a0da4 Clarify known restrictions for model updates. (#40600)
a17e73a0da4 is described below

commit a17e73a0da4710a1fee926ccd2b7abac2ea2ec08
Author: tvalentyn <tv...@users.noreply.github.com>
AuthorDate: Fri Oct 16 12:17:05 2026 -0700

    Clarify known restrictions for model updates. (#40600)
---
 sdks/python/apache_beam/ml/inference/utils.py | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/sdks/python/apache_beam/ml/inference/utils.py b/sdks/python/apache_beam/ml/inference/utils.py
index dfbb01a2904..8cbe836e22d 100644
--- a/sdks/python/apache_beam/ml/inference/utils.py
+++ b/sdks/python/apache_beam/ml/inference/utils.py
@@ -133,14 +133,17 @@ class WatchFilePattern(beam.PTransform):
     1. Any previously used filenames cannot be reused. If a file is added
         or updated to a previously used filename, this transform will ignore
         that update. To trigger a model update, always upload a file with
-        unique name.
-    2. Initially, before the pipeline startup time, WatchFilePattern expects
+        unique name. Do not preserve old modification timestamps when copying
+        or uploading a file (for example, when rolling back to an older model),
+        as files with timestamps older than the pipeline start time are ignored.
+    2. Do not upload more than one model update per polling ``interval``, as
+        matching multiple new files in the same poll can cause undefined
+        ordering or singleton side-input errors.
+    3. Initially, before the pipeline startup time, WatchFilePattern expects
         at least one file present that matches the file_pattern.
-    3. This transform is supported in streaming mode since
+    4. This transform is supported in streaming mode since
         MatchContinuously produces an unbounded source. Running in batch
         mode can lead to undesired results or result in pipeline being stuck.
-
-
     """
     self.file_pattern = file_pattern
     self.interval = interval