[Serve] lazily evaluate autoscaling context#58963
Conversation
Signed-off-by: abrar <abrar@anyscale.com>
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request refactors AutoscalingContext to support lazy evaluation of expensive metrics, which is a great optimization for the controller's CPU usage. The implementation is mostly correct, but I've found a critical bug and some inconsistencies in type hints that should be addressed.
Specifically:
- There's a
TypeErrorin thetotal_running_requestsproperty due to calling properties as methods. - The type hints for some
__init__parameters and property setters inAutoscalingContextare inconsistent with their actual usage, which can lead to confusion and issues with static analysis tools.
I've left specific comments with suggestions to fix these issues. Once they are addressed, this PR should be good to go.
Signed-off-by: abrar <abrar@anyscale.com>
| def total_running_requests(self) -> float: | ||
| # NOTE: for non additive aggregation functions, total_running_requests is not | ||
| # accurate, consider this is a approximation. | ||
| return self.total_num_requests - self.total_queued_requests |
There was a problem hiding this comment.
Bug: Null handling missing in total_running_requests property
The total_running_requests property computes total_num_requests - total_queued_requests, but total_queued_requests can be None (as indicated by the Optional type in __init__ at line 65). When total_queued_requests is None, accessing total_running_requests will raise a TypeError attempting to subtract None from a float. The old dataclass implementation explicitly allowed None for total_running_requests, but the new computed property doesn't handle this case.
Signed-off-by: abrar <abrar@anyscale.com>
autoscaling context need expensive function evaluation, not all autoscaling policies need the data. Lazily evaluate them to save controller CPU --------- Signed-off-by: abrar <abrar@anyscale.com>
autoscaling context need expensive function evaluation, not all autoscaling policies need the data. Lazily evaluate them to save controller CPU --------- Signed-off-by: abrar <abrar@anyscale.com> Signed-off-by: peterxcli <peterxcli@gmail.com>
2 participants
autoscaling context need expensive function evaluation, not all autoscaling policies need the data. Lazily evaluate them to save controller CPU