The document discusses strategies for getting data from streaming data sources into an ERP system. It notes that data streams can include sensors, social media, and internet sources. The data must be cleansed, enriched, and managed before being delivered to the ERP. Events identified from the data streams need to be prioritized and aggregated to avoid overloading the ERP system. Characteristics of the events like randomness, statistical patterns, and importance over time are discussed. The document recommends decoupling the data supply chain from the ERP and using a prioritized event queue to protect the ERP from information overload. Ongoing analysis of data is needed to identify the right events to forward to the ERP and ensure the patterns defining events
Unlocking the Future of AI Agents with Large Language Models
Datastreams into erp
1. GETTING
INFORMATION
FROM DATA
STREAM SOURCES
INTO ERP
SAP Inside Track Copenhagen
By Søren Amdi Bach
Principal Application Architect
KMD A/S
May the 4th. 2018
Who is Søren:
Principal Application Architect at KMD A/S
Professional, curious and enthusiastic technology nerd with a sense of
business and social skills
Some SAP technology words from the recent years:
S/4HANA (Public Cloud, On Premise, Conversion), SAP Cloud Platform,
SAP Leonardo, SAP Solution Manager, HANA Database, Security,
HANA/ABAP: Development, governance, quality, DevOps etc.
Various architect roles the last + 15 years, mainly SAP for the last 12-
13 years
Origin in development/technology: SAP, Microsoft, IBM MVS, Open
Source and Unix
2. _ Internal - KMD A/S
2
INTELLIGENT ERP, INDUSTRY 4.0 … SOMETHING WITH DATA AS NEW OIL
Data Streams
ERPMagic
The simplified (commercial) perspective
A World
with
wi-fi
All the knowledge
you ever need - to
run your enterprise
fully automated
3. _ Internal - KMD A/S
DATA(STREAM) SUPPLY CHAIN
Collection
• Sensors
• Information
scanning
• SoMe
• Internet
Cleansing
• Remove
information
of no interest
• Normalizing
of collected
values
Enrichment
• Aggregation
• Time series
processing
• Correlation
with other
data sources
Management
• Monitor
source health
• Store/archive
selected
information
for further
usage
Deliver
• Analysis
• Identification
of Events
• Notification
to business
processes
Internet Of Things (IOT) Platform
Social Media Information Scanning frameworks
Other Data Stream frameworks
Increasing Entropy in the information
4. _ Internal - KMD A/S
4
▪ Scalable (Cloud resources)
▪ Fast innovation mindset
▪ “Microservice” architecture
▪ Exploratory / Experimenting
▪ DevOps Continuous-Integration
and -Deployment
▪ Limited and Expensive Scalability
(on-prem resources)
▪ Improving / Stability mindset
▪ Monolith (shared Database)
▪ Predictable world
▪ Classical big releases with joint
phases
FROM DATA STREAM TO ERP
Data Supply Chain
ERP
Events
Larger amount of
events emitted in short
time span might
compromise the
performance of the
receiving ERP
5. _ Internal - KMD A/S
PREFERRED CHARACTERISTICS OF EVENTS PROVIDED FOR THE ERP SYSTEM
5
How-to avoid overloading the ERP system
ERP is the critical (less scalable) resource.
• The ERP system should only receive events identifying a unique business event
• An event should trigger update of a defined set business object in the ERP system.
• Only push events that is necessary, of relevance and make sense to the ERP system
Don’t disturb
unnecessary
• Events for the ERP system should be prioritized allowing most important Event types to
be processed at first
• If more events are provided than the ERP system can consume the Events should be put
in a queue
Respect Urgency and
await turn
• Protect against bursts of Events with the same business semantics
• Await sending the Events until the existence of the Business Event is certain
Aggregate Events
before sending to ERP
6. _ Internal - KMD A/S
LIKELY CHARACTERISTICS OF EVENTS IDENTIFIED FROM DATA STREAMS
6
Event is an event by natureRandom
•Events is not in general identified/emitted in a nice well-known predictable manner
Pattern on time series dataStatistical
•A event indicate something in the real-world with a certain probability
•A event might be a false prediction
Importance of a event might be a function of timeImportance f(t)
•Importance of a identified event might varnish over time
•Timestamp on event identification might be on importance
Possible requiring identification of the same eventRepeated
•Identification of events with the same real-world meaning are likely to be repeated
•Hysteresis in the detection might be required to ensure certify of the business meaning of the event
Pattern of events might have a additional business semanticUnknown Patterns
•The identified events might contain new undiscovered knowledge
•Requires some domain knowledge on the real-world topic
7. _ Internal - KMD A/S
▪ Decouple the information for the ERP system
▪ The ERP system load should control the consumption speed of events from the data streams
▪ Consider some sort of prioritized event queue between Data Supply Chain and ERP system
HOW TO – PROTECT THE ERP SYSTEM
7
Data Supply Chain
ERP
11 1223
ERP System is still the bottleneck, but is protected against information overload
8. _ Internal - KMD A/S
8
HOW TO – ENSURE RIGHT EVENTS FORWARDED TO ERP
No quick fix - to identify Events
▪ Requires domain and data science knowledge
▪ Analysis on collected “live data” streams to identify patterns (supervised learning)
Initial go-live set is a proposal
• The initial go-live will be based on a priori knowledge and analysis of existing collected
data
Cotinus validation and retrospective analysis
▪ Non detected real-world events of importance requires analysis of data to search for
correlated event patterns
▪ Data patterns defining an event might change over time
▪ Analysis of larger set of identified events to identify new knowledge (new patterns with
business semantics)
9. _ Internal - KMD A/S
9
DATA SCIENCE PROJECTS AS
INSPIRATION
▪ The identifications of the right events has
similarities to a data science project
▪ The Cross-industry standard process for data
mining (CRISP-DM)
▪ Off-line analysis part (model determination)
▪ On-line/real-time usage of model
▪ Consider methods like
▪ Classification
▪ Segmentation or Clustering
▪ Link analysis
▪ Regression
▪ Time Series Analysis