Parallel model deployment for artificial intelligence using a primary storage system

1. A system, comprising:

a processor;

a memory;

a primary storage system configured to store a training data set and a validation data set;

a data ingester configured to receive raw data for processing through a computational model;

a model trainer configured to:

apply an artificial intelligence framework to the training data set in the primary storage system to generate a model instance of the computational model;

apply the validation data set to measure a reliability value of the model instance;

initiate, responsive to the reliability value not meeting a model validity threshold, retraining of the computational model; and

store, responsive to the reliability value meeting the model validity threshold, the model instance of the computational model in the primary storage system;

a model server configured to deploy the model instance of the computational model for use by a client system; and

a notifier configured to:

determine a new data element received by the data ingester; and

initiate, in parallel with the primary storage system storing the new data element, the model server to evaluate the new data element using the model instance of the computational model.

2. The system of claim 1 , further comprising:

a graphical processing unit, wherein the model trainer is instantiated in the graphical processing unit for operation on the training data set in the primary storage system.

3. The system of claim 1 , further comprising:

a data preparer configured to:

receive the raw data from the data ingester;

prepare the training data set from aggregate raw data; and

store the training data set to the primary storage system.

4. The system of claim 1 , wherein the model trainer is further configured to:

determine an addressable storage location in the primary storage system configured to store the training data set; and

select, using metadata for a plurality of training data elements stored in the addressable storage location, the training data set from the addressable storage location.

5. The system of claim 1 , wherein the primary storage system is further configured to store aggregate raw data.

6. The system of claim 1 , further comprising:

a secondary storage system configured to archive archival model instances, wherein:

the model trainer is further configured to generate a plurality of model instances over time; and

the primary storage system is further configured to:

store active model instances for use by the model server; and

send archival model instances and related training data sets to the secondary storage system.

7. The system of claim 1 , further comprising:

an inference evaluator configured to:

evaluate an inference score for the new data element, wherein the model server is further configured to determine the inference score based on the model instance of the computational model; and

initiate, responsive to the inference score not meeting the model validity threshold, retraining by the model trainer using a new training data set that includes the new data element.

8. The system of claim 7 , wherein:

the model server is further configured to generate metadata for the new data element based on the model instance of the computational model; and

the inference evaluator is further configured to initiate, responsive to the inference score meeting the model validity threshold, post-processing of the metadata.

9. The system of claim 8 , further comprising:

a metadata store configured to aggregate the metadata generated by the model server; and

at least one application configured to use the metadata store to analyze aggregate application of the computational model.

10. The system of claim 1 , wherein:

the primary storage system is configured as a cloud storage system including rack-scale all-flash storage;

the data ingester is further configured to store the raw data into at least one data object stored in the cloud storage system; and

the notifier is configured as a service running on the cloud storage system.

11. A method, comprising:

receiving, by a primary storage system, operations on a training data set from a model trainer;

sending, by the primary storage system, a model instance of a computational model to a model server;

determining that a new data element is received by a data ingester;

initiating, in parallel with the primary storage system storing the new data element, the model server to evaluate the new data element using the model instance of the computational model;

determining, by the model server, an inference score based on the model instance of the computational model;

evaluating the inference score for the new data element; and

initiating, responsive to the inference score not meeting a model validity threshold, retraining by the model trainer using a new training data set that includes the new data element.

12. The method of claim 11 , wherein:

the model trainer is instantiated in a graphical processing unit for generating the model instance using the training data set; and

the training data set is stored in the primary storage system.

13. The method of claim 11 , further comprising:

receiving, by a data preparer, raw data from the data ingester;

preparing, by the data preparer, the training data set from aggregate raw data; and

storing, in the primary storage system, the training data set.

14. The method of claim 11 , further comprising:

applying, by the model trainer, an artificial intelligence framework to the training data set;

generating, by the model trainer, the model instance of the computational model;

applying, by the model trainer, a validation data set stored in the primary storage system to measure a reliability value of the model instance;

initiating, responsive to the reliability value not meeting the model validity threshold, retraining of the computational model; and

storing, responsive to the reliability value meeting the model validity threshold, the model instance of the computational model in the primary storage system.

15. The method of claim 11 , further comprising:

storing, in the primary storage system, aggregate raw data from the data ingester.

16. The method of claim 11 , further comprising:

receiving, from the model trainer, a plurality of model instances over time;

storing, in the primary storage system, active model instances for use by the model server; and

sending, by the primary storage system, archival model instances and related training data sets to a secondary storage system.

17. The method of claim 11 , further comprising:

determining, responsive to the inference score not meeting the model validity threshold, the new training data set that includes the new data element;

applying, by the model trainer and responsive to initiating retraining, an artificial intelligence framework to the new training data set;

generating, by the model trainer and based on the new training data set, a new model instance of the computational model; and

storing the new model instance of the computational model in the primary storage system.

18. The method of claim 17 , further comprising:

generating, by the model server, metadata for the new data element based on the model instance of the computational model; and

initiating, responsive to the inference score meeting the model validity threshold, post-processing of the metadata.

19. The method of claim 18 , further comprising:

aggregating the metadata generated by the model server in a metadata store; and

using the metadata store to analyze aggregate application of the computational model.

20. A system comprising:

a processor;

a memory;

a primary storage system configured to store:

raw data;

a training data set; and

a model instance of a computational model;

means for receiving the raw data for processing through the computational model;

means for generating the model instance of the computational model based on the training data set;

means for deploying the model instance of the computational model for use by a client system;

means for determining a new data element is received for the primary storage system;

means for initiating, in parallel with the primary storage system storing the new data element, the means for deploying the model instance to evaluate the new data element using the model instance of the computational model;

means for determining an inference score based on the model instance of the computational model;

means for evaluating the inference score for the new data element; and

means for initiating, responsive to the inference score not meeting a model validity threshold, retraining of the computational model using a new training data set that includes the new data element.