1. Analytics as a Service (for service assurance with closed loop actions)
Requirement/architecture owner: Srinivasa Addepalli
Source: Edge Automation WG
Thanks to Ramki, Raghu, Margaret, Chaker, Vijay, Jack, Tom, Donald and Frank for their mindshare, feedback and/or contributions
Disclaimer: It is WIP and some details will change as we develop more understanding on integration aspects with DCAE and PNDA

2. Big Picture – Analytics
Reactive
Acted on escalation, productivity impact in hours
Pro-active
Continuous monitoring, act on alerts, little impact on end users, but may miss alerts due to volumes
Predictive
Identifies and then attempts to prevent SLA-impacting events
Need AI (ML/DL) analytics
Need Big data framework

3. Traffic Classifica-tion
Prediction Examples
Intrusion Detection
Traffic Classifica-tion
Anomaly Detection
Fault Prediction
Path Prediction
Resource Prediction
Traffic Prediction
Others

4. Network Analytics – Current method & Challenges
One (or very few) analytics framework deployments
All data is fed to central deployment
Only VNF and application analytics
Not Edge/5G friendly
Scalability challenge
Lack of faster closed loop control
Big Data analytics framework deployment is hard
2 weeks to deploy - Think about multiple deployments
3 to 4 weeks to configure and tune
Continuous monitoring is a challenge
Not leveraging hardware accelerators such as ML/DL algorithm accelerators & math operations
Not leveraging infrastructure metrics (that help in accuracy of prediction)

5. Big Data Analytics Stages
Distribution
Store Models DB
Collection Agents
Data Storage (Data Lake)
Training
Inferencing
Action Correlation
Models
Action Controllers (External controllers such as Kubernetes*, APPC, SDNC)
Training apps
Inferencing apps
Correlation apps
*Other names and brands may be claimed as the property of others.

6. Data Analytics Framework
Needs / Requirements

7. Stack Goals – Decentralization
Cloud Site
Central Site
Collect
Distribute
Store
Train
Predict
Correlate
Action
CENTRAL ANALYTICS
Edge Clouds
Private Cloud
Public Cloud
Predict
Train
Train
Action
Store
Store
Small Edge
Medium Edge
Large Edge
Collect
Collect
Collect
Distribute
Distribute
Distribute
Predict
Store, Train
Correlate
Predict(s)
Action
Action(s)
Challenges:
Petabytes of data need to be transferred
Not edge friendly
Data Sovereignty challenges

8. Stack Goals – Bulk Deployment & Configuration
Automation & Configuration
Edges could be in thousands
Each edge would have its own site orchestrator (e.g., Kubernetes*)
Need for centralized automation
To deploy various parts of analytics stack in multiple locations.
Configure them to make them work together
Support for dynamic edges
Goal: Deploy and configure big data analytics framework in hours instead of months
Medium Edge
Predict
Large Edge
Collect
Distribute
Train(s)
Predict(s)
Action(s)
Regional
Small Edge
Action
Train
Store
*Other names and brands may be claimed as the property of others.

9. Stack Goals – Use Known Big Data Frameworks
Goal is to leverage known and open source packages
Choose right versions
AI (ML/DL) training and inferencing
Make them work together
Gaps and fixes
*Other names and brands may be claimed as the property of others.

10. Stack Goals – Follow Cloud Native and Micro-Services Design Patterns
SECURITY
ISTIO* CA, Envoy* proxy for Mutual TLS among PODs.
ISTIO ingress for communication outside of clusters.
ISTIO RBAC
 Security away from applications
LOAD BALANCING & CI/CD
ISTIO Envoy with Cilium acceleration
Visualization and monitoring
A/B testing & Canary
OPERATORS
To bring up PODs.
To configure using CRDs.
FUNCTIONS
To enable developer functions to get executed in the pipeline.
Native
*Other names and brands may be claimed as the property of others.

11. Stack Goals – Reduce training time
Distributed Learning with Data Parallelism
Leverage Hardware Accelerators for Performance
Leverage NI optimized software libraries

12. Stack Goals – Federated/Distributed Learning
Distributed learning orchestration
To honor data sovereignty
Data does not leave the site
Aggregation server based model
TensorFlow* based DL
Aggregator
Data Owner1
train
Data Owner2
*Other names and brands may be claimed as the property of others.

13. Data Analytics Framework
An opinionated stack

14. Deployment and Automation using ONAP4K8S
Stack – Packages & Enhancements to Realize Analytics Stack – Release 6 goals
An opinionated stack
Helm Charts
Deployed using K8S in each site
R6 time frame goals
Container images – Build from source
Few Customized Day0 profiles
Day 2 templates for dynamic configuration
Security via ISTIO*
Deploy with ONAP- K8S.
NATS
Sample applications
Code: demo/tree/master/vnfs/ DAaaS
Deployment and Automation using ONAP4K8S
Collection Agents
Distribution
Data Storage (Data Lake)
Training
Store Models DB
CollectD
Node Exporter
C Advisor
Prometheus*
Operator
HDFS
M3DB, M3DB operator
Spark*, Spark operator
Scikit learn
Horovod
TF, TF Operator
Model Dispatcher
Minio*
Minio operator
Collected Operator
HDFS Writer
HDFS Operator
Inferencing
Action Correlation
Action Controllers
TF and TF Serving operator
kNative
ML Serving & Operator
Messaging
Visualization
NATS
Kafka* Broker
Kafka Operator
Zookeeper*
Grafana*
Hue*
*Other names and brands may be claimed as the property of others.

15. Roadmap

16. Roadmap / Wishilist
Uniform OS across all containers ( ClearLinux or Ubuntu )
Build from source code (Leverage instruction set optimized libraries, Enable security patching etc..) – Get control of container SW packages.
Automation – One click deployment of stack in multiple locations
Pipelining and DAC support for analytics applications
Monitoring – Stack as well as analytics applications
Native support – for action functions
Federated Learning – to maintain data sovereignty challenges
Generic action functions for K8S* and ONAP*
Model LCM and Model security
Leverage hardware accelerators and OpenVINO™
Integration with ONAP DCAE
Inferencing using non-TF (non-DL)
*Other names and brands may be claimed as the property of others.

17. Call to Action FEEDBACK… FEEDBACK… FEEDBACK…
It is open source, contributions are welcome.
Source code repo is in LFN/ONAP*. But… It is modular and can be deployed without ONAP
FEEDBACK… FEEDBACK… FEEDBACK…
Try & let us know.
New collection mechanisms?
Lightweight messaging?
Model repositories?
Any data lakes?
Want to try your data analytics applications? Please contact us.
*Other names and brands may be claimed as the property of others.