This project describes how does the orchestrator interacts with nodes. The protocol is a OpenApi based protocol.
The general architecture of the system is described below.
Any node must implement this protocol to be compatible with the orchestrator.
The outlier_protocol is a simple protocol based on http messages that are grouped into 3 main categroies: health, train and predict. Each group is described below.
Models must implement the /health endpoint. The orchestrator will send a get message at a variable time. If the model does not answer, orchestrator must consider the model as unavailable, and data must not be sent to this model.
With these group of endpoints, it is described how the models will be trained in order to acquire data for disruption prediction.
Protocol begins with a post message to the /train model endpoint that contains the number of discharges that will be sent to the models. Models that want to accept the training, answer with a 200 response. Then the orchestrator sends, one by one, the training discharges. After a discharge is received, models must acknowledge it. When all discharges are sent, models shall start the training.
When the training is done, models must inform the orchestrator, to the /trainingCompleted orchestrator endpoint. Each model will inform about the training stats that has achieved.
Orchestrator will send a post message to the /predict endpoint of every model in order to start a prediction. Models must answer with the predicted value.
