Automated Error Handling for Deployed Models
- Observability
- H2O MLOps
This lesson covers how a deployed model can help identify and handle issues on its own. It shows the monitoring data available for a deployment, including the number of predictions coming in over time, average latency, and whether the model is showing any drift. It then demonstrates sending a data row to the endpoint from the command line and getting back a prediction, followed by sending a row with the wrong data type, in this case text where a number was expected, and getting back a 400 status code along with information pointing to what went wrong, so downstream tasks can build error handling around it.
Key takeaways
- What monitoring data is available for a deployment: prediction volume over time, average latency, and drift
- How sending the wrong data type to a model endpoint returns a 400 status code and details on what went wrong
- Why returning specific error information for bad input lets downstream tasks build their own error handling around it
Start owning your AI