[Info-vax] HPE buys Cray

Grant Taylor gtaylor at tnetconsulting.net
Sat May 18 14:49:29 EDT 2019


On 5/18/19 12:30 PM, Kerry Main wrote:
> Interesting product offering at HPE - using AI and "machine learning" 
> methods to proactively manage Systems/Applications i.e. identify issues 
> before they impact the business.
I feel like much of the AI / ML is noise if not hype of said noise.

$LastJob tried adding IBM Watson to system administration.  They ran 
into a number of problems doing so.  First, the amount of noise made the 
quality of the signal so low that it wasn't worth wasting time on. 
Second, the systems weren't nearly as consistent (cattle) as people 
thought they were.  In fact, they were quite unique (pets) on many 
different levels.  This made it quit difficult to get anything of value 
out of the AI / ML.

What they did get out of the AI / ML was largely already known.  Watch 
all your logs.

10  Look for anything that's unknown.
        If you're starting at zero, everything is unknown, and this is okay.
20  Simply pick the most common thing and decide if it's acceptable or not.
30  Then filter /that/ thing out of the stream of data to process.
40  GOTO 10

As you iterate through the above process you will learn more about your 
environment, what is normal, what is abnormal, and what warrants 
immediate action.

The thing that AI / ML did do was to find some things that were many 
Many MANY iterations deep.  Some of which did have a cross interaction 
with each other.  In hindsight, they made obvious sense.  But they were 
of relatively low value, especially compared to the other more 
significant things that the humans saw using traditional methods.

> I do know one large site I was at was looking at adapting Splunk to do 
> big data style analysis of all the infrastructure (servers, storage, 
> network) / application / appliance log files in their IT environment.

I think that taking raw data streams and hiding known good signals to 
the point that only unknown and / or bad signals show up is a decent 
idea.  But I don't think that AI or ML is necessary.  Especially when 
what is functionally a filter can do much of what is needed.



-- 
Grant. . . .
unix || die



More information about the Info-vax mailing list