seldon-core
seldon-core copied to clipboard
SeldonIO
An MLOps framework to package, deploy, monitor and manage thousands of production machine learning models
## Describe the bug ## To reproduce ``` >>> import json >>> import seldon_core.proto.prediction_pb2 >>> from google.protobuf.json_format import MessageToJson >>> msg = seldon_core.proto.prediction_pb2.Metric( ... key="my_metric", ... type="GAUGE", ... value=0, ......
## Describe the bug When deploying the seldon controller manager with a different name than `seldon-controller-manager`, the deployment fails due to this name being hardcoded [here](https://github.com/SeldonIO/seldon-core/blob/master/operator/constants/constants.go#L44). This is an issue...
The current HPA spec we allow is at v2beta2 and does not allow [Scaling Behaviour](https://kubernetes.io/docs/reference/generated/kubernetes-api/v1.22/#horizontalpodautoscalerbehavior-v2beta2-autoscaling) to be specified. Need to investigate if * We upgrade to v1 spec * Allow...
# Custom Error Doesn't work for MLFlow Server ## Describe the bug When defining custom error handling according to documentation and using an MLFlowServer the error handler doesn't work. ##...
Annotations allow for advanced scaling using external metrics from the Kube Metrics Adapter (KMA). For example, querying Prometheus to calculate a metric to scale on (i.e. Requests per second per...
Currently, the executor and the language wrapper use the OpenTracing libraries for tracing. However, that project is now moving into OpenTelemetry. OpenTelemetry claims to offer a standard protocol that works...
## Describe the bug Placing a metadata.yaml file with metadata information about the model in the model s3 bucket does not work when using the prepackaged tensorflow server and the...
Several users have reported issues when running TF models / Pytorrch models in the python wrapper due to the gunicorn multiprocess. This is a known issue and has formal fixes,...
Currently Ambassador support is implemented by annotations on the predictor services. For more flexibility and better integration switching to the CRDs would be preferable.
Currently the fixed `drift_batch_size` set by the user can result in batches bigger than the threshold to be fed through the detector: https://github.com/SeldonIO/seldon-core/blob/master/components/alibi-detect-server/adserver/cd_model.py#L89 The default behaviour should be that if...
