⭐Scheduling a streaming inference
After you create a model, you can use it to monitor your asset in streaming. To use your model to monitor your asset, you do the following.
1. Before scheduling
:white_check_mark:The data being used to inference the model: This inference dataset should have the same schema and data granularity as the training data set for the model.
2. Schedule a streaming inference
When the model status is completed, click the button to schedule a streaming inference to start collecting data and detecting anomalies.
3. Set alert configuration
Set alert configuration, specify alert name and sensitivity, and alert correlation and suppression. 
Select or create new hooks for the alert.
| Parameter | Description |
|---|---|
| Alert Correlation and Suppression | Correlated anomalies if they occur within [x] data points of each other. Send one alert for correlated anomalies. Send an alert for each anomaly: get notifications for new anomalies by receiving alerts at hooks. |
| Sensitivity | From 1 ~ 100, to define detector's sensitivity. The bigger the number the more sensitive to anomalies. Set a lower sensitivity to get notified only when severe anomalies are detected. Set a higher sensitivity to detect anomalies that are less severe. |
4. Streaming Inference parameters
Specify the name for the inference and the time period that your model performs inference on the data coming from your pipeline.
` Inference start time`: This start time can't be in the past. The first timestamp equal to or greater than the start time given will be used. Your data source must have data at the specified start time and the number of data points available must equal to or greater than your training sliding window (model_name's sliding window = xxx).
Data delay Offset (this is an optional field): the amount of time (in seconds) you expect the data to be delayed for inference. For example, your source data comes every 5 minutes, so by default (i.e., data delay offset In seconds = 0) the inference schedule assumes that records with a timestamp of 01:30:00 will be ready for inference by 01:35:00, records with a timestamp of 01:35:00 will be ready at 01:40:00, and so on. If you expect a 10-minute data delay (i.e., data delay offset In seconds = 600), then the scheduler will inference records with a timestamp of 01:30:00 at 01:45:00, inference records with a timestamp of 01:35:00 at 01:50:00, and so on.
:warning:Metrics Advisor for Equipment uses UTC time that specifies when to start running the inference to detect anomalies.
For the dataset, select or create an inference dataset.
5. Update a paused inference
If users already created a streaming inference that is running in other models, users could replace the trained model with a paused inference. Previous data will be kept, and anomalies that were detected with the old model will still be visible next to newly detected anomalies in the same inference data. Updating the model is only recommended if there weren't any significant differences in the training parameters.
:warning: If an inference isn't listed here, make sure to pause it first. The inference will restart and be streaming immediately after updating.
6. Set a replay
After a model is doing streaming inference, users could choose to schedule a replay. This will trigger a backfill on your selected timestamp immediately to fix a failed inference or to override the existing data.
Replay: overwrite the data in a historical time range by running the inference with its current settings.
A max of 2000 data points can be replayed. For example, with a data frequency interval of 1 minute, you can replay a time range of about 1 day.
Schedule replay/backfill for an existing streaming inference. Select start date and time and Select end date and time.
The current replay must finish before creating a new one. Wait until the evaluation completes before creating a replay. Replay is re-triggered only on the selected range.
7. Inference detail
Once start the inference process. Click the model name, and then the Inference detail tab allows you to review inference details such as the inference name, the dataset, and inference start time.
There are three kinds of inference status.
| status | Description |
|---|---|
| Active | Streaming Inference is in processing. |
| Inactive | Streaming Inference is paused by users or Inference is failed for some reason. If Inference is paused by users, you can restart inference. |
There are three kinds of actions for streaming inference.
- Pause
This will halt the inference process. You could stop the streaming inference from destroying incurring costs. From the Model details, under the inference tab, choose restart.
- Restart
This will resume a stopped inference schedule. From the Model details, under the inference tab, choose Pause.
- Deleting a paused inference
When you don't have any more use for your streaming inference, you can delete a stopped streaming inference by choosing Delete.
You can click the View details button for the error messages when you find the training inference status gets to Inactive.