FIWARE ML as a Service POC

A simple Proof Of Concept for a FIWARE MLaaS.

Concept

This repository contains the code for a conceptual 'Machine Learning as a Service' component providing prediction service from/to a FIWARE context broker. In a nutshell a MLaaS can be used by a Consumer Application using the following procedure (also described as a sequence diagram below):

• As a prerequisite (flow 2-4), a MLaaS (MLModel) has been deployed and has subscribed to the Context Broker in order to be notified when a consumer application request to use its service. The notification contains a query for the MLaaS to identify the data on which the prediction shall be made.
• A consumer application creates a MLProcessing entity in the context broker (flow 5), which contains a reference to the desired MLaaS and a reference (as a query) to the data on which predictions have to be made.
• Finally the consumer application subscribes to the prediction (data results of the prediction performed by MLaaS)
• The creation of the MLProcessing entity by the consumer app triggers a notification towards MLaaS. The MLaaS uses the content of the notification to find out which data to use for predicting and susbcribe to changes of that data (flow 7-8).
• Whenever that data changes, a notification is received by MLaaS and triggers a prediction (flow 9-10).
• MLaaS saves the prediction into the Context Broker (flow 11)
• Finally the consumer application receives a notification of the predicted data (flow 12).

Getting started

• Configure the global variables in ctes.py

• Create the MLModel entities (adapt the variables in the file)

python3 create_mlmodel.py

• List your models to check the creations went fine

python3 get_mlmodels.py | jq

• For tests purposes, you can manually create a WaterConsumption entity that will hold input data and prediction results

python3 create_waterconsumption.py

• Add a relationship between the WaterConsumption entity and a previously created MLModel entity

python3 add_measure_waterconsumption.py

Similarly, you can use the delete_relationship_to_mlmodels.py to delete a relationship between a WaterConsumtion entity and a MLModel entity.

• Create a subscription for the BentoML proxy to be notify when there is an update on a WaterConsumption entity

python3 create_subscription.py

• Update a value on the WaterConsumption entity to trigger a new prediction in the BentoML module

python3 add_measure_waterconsumption.py

• Update a value in a MLModel entity (e.g. the URL of the BentoML module)

python3 update_mlmodel_attribute.py

Building a Docker images

• Build a new image

docker build -t easyglobalmarket/bentoml_ngsild_proxy .

• Run it

docker run -p 5000:5000 easyglobalmarket/bentoml_ngsild_proxy