feat: SDK support for model monitoring#1249
Conversation
6481291 to
68b6fa4
Compare
…nts; batch prediction use case will be implemented separately)"
a154c08 to
327611c
Compare
99987d3 to
ee05588
Compare
…on-aiplatform into model-monitoring
|
|
||
|
|
||
| class RandomSampleConfig(_SamplingStrategy): | ||
| def __init__(self, sample_rate: Optional[float] = 1): |
There was a problem hiding this comment.
Please change this or clarify the behavior when sample_rate is None.
| return ( | ||
| gca_model_deployment_monitoring_job.ModelDeploymentMonitoringScheduleConfig( | ||
| monitor_interval=duration_pb2.Duration( | ||
| seconds=self.monitor_interval * 3600 |
There was a problem hiding this comment.
This conversion can be surprising. If the original schedule config (defined in the service protocol) expresses this in seconds, why not keep using seconds (and rename the variable as something like monitor_interval_seconds) instead of int hours? How will the user express "every 10 minutes" using the ScheduleConfig class here?
There was a problem hiding this comment.
model monitoring only supports hourly schedules. even if the user specifies something like 1.6, it'll be rounded up to the next hour
There was a problem hiding this comment.
so under the original protocol, even if the user passes something like seconds = 3500, it'll get rounded up to 3600 behind the scenes
There was a problem hiding this comment.
Given the current behavior of the service (rounding to hours), this makes sense. Do we know that the service will not be updated to support more fine-grained monitor interval? If that happens do we have an easy path of updating the library to support that?
There was a problem hiding this comment.
I have sync'd with @qijing93 offline and there's no additional support planned for fine-grained monitor intervals.
…on-aiplatform into model-monitoring
…on-aiplatform into model-monitoring
…on-aiplatform into model-monitoring
This patch only adds model monitoring implementation for models deployed to an endpoint. The batch prediction use case will be addressed separately in future PRs.
To-do list:
Fixes b/231988321 🦕