Machine Learning: 2 Books in 1: Machine Learning for Beginners, Machine Learning Mathematics. An Introduction Guide to Understand Data Science Through the Business Application

8. 
Model Scoring
The terms "Model Scoring" and "Model Serving" are being used
synonymously throughout the industry. Model Scoring" can be defined as
"the process of generating new values with a given model and new input
data". Instead of the term "prediction", a generic term "score" is being used
to account for the distinct values it may lead to, as listed below:
List of product recommendations.
Numerical values, for the "time series models" as well as
"regression models".
A "probability value" that indicates the probability that a new
input can be added to a category that already exists in the model.
An alphabetical value indicating the name of a category that most
closely resembles the new input data.
A "predicted class or outcome" can also be listed as a model
score particularly for "classification models".
After the models have been deployed they can be utilized to score based on
the feature data supplied by the prior pipelines or provided directly by a
"client service". The models must generate predictions with the same
accuracy and high performance, in both the online and the offline mode.

---

“Offline”
The "scoring service" would be optimized in the offline layer for a big
volume of data to achieve high performance and generate "fire and forget"
predictions. A model can send an "asynchronous request" to initiate its
scoring, however, it must wait for the batch scoring process to be completed
and gain access to the batch scoring results before the scoring can be
started. The "scoring service" will prepare the data, produce the features as
well as retrieve additional features from the "Feature Data Store". The
results of the scoring will be stored in the "Score Data Store", once the
scoring has been completed. The "broker" will be informed of the
completion of the scoring by receiving a notification from the service. This
event is detected by the model, which will then proceed to collect the
scoring results.