
Peter Burris101

















































Rob Callery
Does this mean a build out of infrastructure at the Edge?
John DesJardins
you can't do predictive analytics AT the edge without algorithms and models. These need to be built using ALL the data, centrally.

Brian Patrick Murphy
the volume of machine data will significantly restrict the ability to transmit to data centers

Sean Riley
data transmission costs and ability will be central to the success of IoT. This will be important on a production line, where currently a small percentage of data is actually used for analysis, but will be more important to field services.

Stuart Miniman
the last mile has always been challenging for networks - #IoT will see this even more. Decisions need to happen on the edge and intelligence as to what comes back to central location.
John DesJardins
data from edge doesn't need to reach data center or cloud in real-time, but it DOES need to reach there for doing machine learning and predictive model training.

Eric Kraemer
From an Industrial perspective, most telemetry today runs at 1Hz or lower frequency. Moving to advanced analytic applications that require machine data at 10's to 100's Hz will certainly be cost prohibitive to centralize

Andrew Foster
Edge offers many benefits including Lower latency, Better determinism, Increased security, Lower data comm costs

Calvin21
Entirely agree. Key is in determining what data at the edge is perishable (only valuable for a limited period of time, and only at the edge), and what data is actually worth storing/analyzing in a historical or trend-based context.

Edgard Capdevielle
Agree! In industrial installations there are use cases were cloud can play a big role (process analytics) but most use cases still have requirements for edge processing.
John DesJardins
@ejkraemer you can put real-time analytics at edge, but you still eventually need data to be central to train models.

John Furrier
The question that I have is "how do edge #IoT device" be agile and intelligent and deal with multiple clouds? Doesn't latency kill in this new model

ClearSky Data
More businesses will turn to colocation centers in metro areas

Sean Riley
You are correct. however, not all data needs to be used to build very accurate predictive models so edge analytics are used to generate actions (alerts) and filter out data that isn't useful for model development or CI efforts.

yaron haviv
move to edge analytics and distributed data

David Floyer
A key characteristic of most IoT is machine to machine communication at millisecond speed - 95%-99% of data will stay and die at the edge or close edge. Key data will be uploaded for cloud support and cloud training of edge software.

William Geller
I agree with this. Edge gateways and hubs will become more powerful along compute and storage, and leverage a decentralized on-prem infrastructure, with only key data assets transferred to the Cloud (or on-prem core) for further analysis.

Eric Kraemer
It can be important to differentiate between decision making, analysis, and data management. There are important distinctions

John Furrier
@yaronhaviv does edge analytics & distributed data mean new software?
John DesJardins
@1seanriley if you want to model time series data, you will need to reconstruct your time series in your training data. This doesn't mean that the information shouldn't be compressed and also can be transferred in batch

Bert Latamore
Totally agree. I see network capacity as a limiting issue. If. for instance, a commercial airliner generates 1TB of data with each flight, multiplied by the number of airliners in the air each data that comes to multi Pbytes of data from airlines

Kevin Terwilliger
We had a partner look at the cost of passing all data to the cloud for IoT deployments. For a simple sensor module (~15 sensors) it cost ~$800 a month just to ingest & store the data.. that will hurt when you turn on a factory of fleet of sensors

Calvin21
Can anyone find me an unfunded edge analytics startup? This is a huge growth area in industry...winners will help to determine perishable vs. long-term valuable data, vs. don't capture and/or don't store data
John DesJardins
I wonder how many people here are industrial vs. infrastructure/network background vs. data scientists?

yaron haviv
from data lakes to integrated edge "appliances" stream/db/ML in one

William Geller
@awf1967 The problem has always been compute/storage scalability in an unforgiving environment - manufacturing floors, for example
John DesJardins
I can say with certainty that Autonomous Driving algorithms MUST be built with ALL the data.

Dave Vellante
@yaronhaviv right - bring the model to the data

yaron haviv
where is the edge? Last mile, telco, metro?

William Geller
@BertLatamore Interesting anecdote. Could we begin to see airports turning into edge server hubs, and sell service to major airlines?

Eric Kraemer
Motion & acoustic modeling - for good accuracy we should be considering sample rates in the 10's of Hz at least. 10-100 samples a second per tag. And you can't just burn a model in and walk away. Conditions in the real world change
John DesJardins
as we move towards more "AUTONOMOUS" analytics at the Edge, building the logic will require data also to reach data scientists and Cloud or Data Center.

John Furrier
John DesJardins does timeseries data matter in real time data when the time is changing so the series analysis also changes? thoughts

yaron haviv
greater focus on density, power,

Andrew Foster
@williamgeller new generation of ruggedized Smart Edge gateways and Fog nodes can help
John DesJardins
Also think about CONTEXT - Context is missing at the Edge - Weather, Customer, Product, and other enterprise data

Vidcentum R&D
Yes.
John DesJardins
Time-series data IS the number one type of data at the Edge

Kevin Terwilliger
@awf1967 all great points! a lot of people overlook the improved security. First it is more secure because you don't have to move everything and second because the intelligence at the edge can make up for less intelligent things including analyzing risks
John DesJardins
Also - SEASONALITY - how do you model for seasonal factors when you only have MINUTES of data or less?

Eric Kraemer
When it comes to machine sensor data - we never have all the data with respect to the physics that are being sampled. We have just that - samples that live within the limits of the sensing device and communications infrastructure.

yaron haviv
serverless is the new stream processing

Kevin Terwilliger
@dfloyer We like to call that data 'perishable' - it is only valuable for a split second
John DesJardins
just because the INSIGHTS are perishable, doesn't mean that the data doesn't also have value for training models.
Ricardo Costa
@kdtwill Value can assume many forms. Yes, it has a short life for alerting and some level of decision making, but it may live longer for research and analysis (e.g. re-train analytical models).

Bert Latamore
I expect that a lot of industrial data will be analyzed in real time at the site for actionable info -- e.g. out of range readings indicating a developing mechanical problem.

Peter Burris
For those wondering, by "basic physics" we're really focusing on the costs of latency and entropy, which are real and won't change due to 5G.

Bert Latamore
The vase majority of edge IIOT data will go directly to tape on site. As David Floyer likes to say, a truck full of tapes has the highest bandwidth. & it is cheap transport.
John DesJardins
NHTSA has also issued directives for retaining all data used by Autonomous Driving systems.