Cosmos - Change Feed Continuation Token Parsing Bug #43018

• Package Name: azure-cosmos
• Package Version: 4.9.0
• Operating System: Debian
• Python Version: 3.12.11

Describe the bug

Using the ContainerProxy.query_items_change_feed with the continuation fails with a ValueError when provided a valid continuation token returned by Cosmos DB (NoSQL mode). Example error:

ValueError: year 57679 is out of range

To Reproduce
Steps to reproduce the behavior:

1. Create a Cosmos NoSQL account, database and container
2. Using the Python SDK, call query_items_change_feed on the container instance with a start_time parameter supplied.
3. Capture the continuation token from client_connection.last_response_headers["etag"]
4. Call query_items_change_feed on the container instance with the continuation parameter set to the value from 3.

Expected behavior
No errors should occur, the change feed query should return expected results, allowing a client to persist the continuation token and pull down further change feed information.

Additional context

This is happening due to bad handling of the continuation token. At azure/cosmos/_change_feed/change_feed_start_from.py:198, the SDK attempts to parse the point in time with the following:

point_in_time = datetime.fromtimestamp(point_in_time_ms).astimezone(timezone.utc)

However, the relevant part of the base64 encoded continuation token has the time present as a Unix timestamp in milliseconds, example extracted from a token:

{"Type": "PointInTime", "PointInTimeMs": 1758011093508}

The from_timestamp method expects a Unix timestamp in seconds. Hence passing the timestamp in milliseconds is causing the ValueError. The variable is called point_in_time_ms which is correct but this is not being used correctly. One possible fix is something like:

point_in_time = datetime.fromtimestamp(point_in_time_ms / 1000).astimezone(timezone.utc) + timedelta(milliseconds=point_in_time_ms % 1000)