►
From YouTube: Webinar: Taking Serverless and AI to the Next Level
Description
The two most trending technologies today are AI and serverless and surprisingly enough, they even go well together to enable agile and intelligent businesses. In this session, Yaron Haviv from iguazio and Tomer Rosenthal from Microsoft will provide an overview of serverless architectures and efforts to encourage collaboration and portability through CNCF working groups.
Yaron and Tomer will present new serverless platforms like nuclio which boost application performance and simplify the adoption of cloud native architectures across multiple cloud and edge locations. They will explain how serverless can accelerate AI from development to production and integrate local or cloud provider event triggers and data sources. Participants will learn how to flexibly test, deploy and auto-scale the same function codes in a public cloud like Azure, on an Edge appliance, or on a developer’s laptop.
Yaron and Tomer will demonstrate these technologies with cool AI use-cases such as sentiment analysis, image classification and predictive maintenance. They will then dive into how to build a fast AI serverless app on Azure with AKS, event-hub and ML/AI APIs, and how to deploy the same functions at the edge using open-source frameworks to address data gravity and latency. You will learn how functions are built from a simple ‘Hello World’ through real-time stream processing, deep-learning and multi-stage workflows and how to distribute or update AI models in online systems, all with no infrastructure setup and only minimal coding.
Join us for KubeCon + CloudNativeCon in Barcelona May 20 - 23, Shanghai June 24 - 26, and San Diego November 18 - 21! Learn more at https://kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy and all of the other CNCF-hosted projects.
A
Okay,
tell
you
what
let's
go
ahead
and
get
started
again.
Welcome
to
today's
CMC
s,
webinar
taking
serverless
and
a
I
to
the
next
level.
My
name
is
Dario
de
teed
I
am
a
pounder
at
a
company
called
stack
point
class
where
we
made
it
really
easy
for
you
to
deploy
manage
bunch
of
native
tooling
on
the
cloud
provider
of
your
choice,
I'm
also
a
CNCs
ambassador
I
want
to
thank
you
all
for
making
the
time
to
join
us
today.
I
also
want
to
extend
a
special
thanks
to
the
presenters
today
from
aguado
and
Microsoft.
A
We
appreciate
them
sharing
their
expertise
and
on
moving
from
the
server
Lachey
I
and
some
of
the
work
they've
done
around
nucleo
a
little
bit
of
housekeeping
during
the
webinar.
You
will
not
be
able
to
talk
to
them
attendee.
There
is
a
chat
window
that
you'll
see
down
at
the
bottom
or
and
there's
also
a
QA
button
right
next
to
it,
go
ahead
and
click
one
of
those
and.
B
A
And
the
presentations
permit
we'll
go
ahead
and
jump
in
and
present
those
to
the
speakers
with
that,
let's
go
ahead
and
introduce
today's
speakers.
We
have
from
Microsoft
Tomer
Rosenthal,
where
he
works
on
commercial
software
engineering
and
as
many
of
the
commits
to
the
nucleo
project
that
we'll
be
discussing
today.
Additionally,
we
have
your
own
javi,
who
is
the
CTO
at
equal
deal,
and
so
thanks
again
for
making
time
for
us
today
guys
I'll
hand
it
over
to
you.
C
So
hi
everyone
hi
everyone.
So
we'll
start
my
name
is
Ron
I'm,
a
founder
and
CTO
for
a
company
called
agua
CEO,
well,
mainly
be
speaking
about
surveillance
and
nucleo,
or
not
much
about
what
iguazu
cos
on
my
spare
time.
I'm
also
service
in
CF
ambassador
as
well,
and
also
working
on
the
service
working
and
some
standardization
in
the
service
world.
B
Hi
guys,
my
name
is
Timur
my
work
from
Microsoft
in
the
division.
That's
called
commercial
software
engineering.
Essentially
what
we
does.
What
we
do
is
we
contribute
code
to
project
to
open-source
projects
and
we're
working
with
a
lot
of
companies
on
their
challenges,
and
then
we
contribute
the
solutions
to
github
or
any
other
public
eye
rip.
Oh,
and
we
recently
worked
with
the
guys
in
Iguazu
on
nucleo.
C
Great
so
Wow
so
we'll
start.
Hopefully
most
of
people
are
already
here
and
we'll
start
with
a
little
bit
of
background
about
Cerberus
and
cert
the
challenges
around
it
and
the
state
of
the
universe
and
then
we'll
dive
into
multiple
demonstrations
and
use
cases.
How
essentially
surveillance
is
transforming
what
it
is.
We
call
big
data
or
fast
data
and
allows
a
much
more
rapid
development.
C
Now,
if
you're
thinking
about
what
everyone
is
trying
to
do
today
in
the
context
of
data
and
modern
applications,
essentially
being
able
to
grab
a
lot
of
data
and
analyzing
it
and
generate
eventually
some
insights
or
actions,
you
know
if
you
just
analyze
data
for
a
sake
of
fun.
It's
a
it's
a
bit.
Maybe
it's
nice
when
it's
not
productive.
What
you're
trying
to
do
is
take
data
analyze
it
and
generate
actions,
and,
what's
today,
more
and
more
you'll
see
the
sound.
C
So
what's
terminus
here
allows
you
is
this
flexibility
layer
and
this
highly
scalable
and
elastic
layer
of
computation,
and
that
needs
to
run
on
top
of
some
data
services?
It's
not
the
challenge
today,
anymore.
In
the
cloud
you
go
to
any
cloud
provider,
you'll
get
databases
as
a
service
object,
store,
screening
solutions,
even
like
five
of
each
per
provider.
Ok
and
no
shorter
than
in
that-
and
also
within
the
other
solutions-
are
many
commercial
solutions,
including
what
we
do
in
quasi
on
the
data
side
and
what
this
architecture
needs
to
provide.
C
Essentially,
the
first
thing
is
immediately
insights:
the
ability
to
process
a
lot
of
data
fast
and
generate
results.
The
second
is
delivering
smarter
decisions
that
impact
the
business,
so
introducing
AI
include
listening
aggregates
as
an
enrichment
of
different
types
of
data
and
eventually,
if
you
do
all
those
things
and
it
takes
too
long-
that's
also
a
challenge.
So
what
service
and
micro
services
in
general
allow?
C
You
is
rapid
time
to
production,
and
it's
not
just
the
development
portion,
because
people
tend
to
forget
there
is
the
operationalization
phase,
which
is
usually
takes
80%
of
the
time,
which
is
you
know,
deploying
testing
running
in
production,
scaling
out,
etc,
and
this
is
the
big
part.
The
service
solves,
not
necessarily
the
development,
because
it's
still
an
early
phase.
Sometimes
the
development
is
actually
harder,
but
what
service
really
solves
is
the
operational
operationalizing,
your
application
at
scale?
So
that's
a
server
previewing
to
how
service
is
introduced
into
this
space
of
data
and
analytics.
C
You
know
there
is
a
small
question
here.
Is
it
limited
to
the
cloud
and
what
we're
trying
to
show
here
in
this
session
that
you
could
get
nice
things
in
the
cloud
and
that's
will
be
sort
of
the
age
of
peace
and
but
also
you
can
run
in
the
edge
or
on
Prem,
using
open
source
solutions
or
commercial
solutions.
B
C
And
I
think
people
get
into
this
realization
is
that
sometimes
data
locality
does
matter
you're
taking
a
lot
of
video
feeds
and
you
want
to
analyze
them
from
some
street
cameras.
You're,
probably
not
gonna,
shape
them
into
the
cloud
you're
doing
IOT
at
the
factory.
I
will
show
one
use
case
for
that.
You
have
to
process
out
the
data
at
the
factory.
C
So
when
talking
about
services,
there
are
a
bunch
of
issues
and
good
things
and
bad
things.
So
the
good
thing
is
it's
pretty
cool
everyone
likes
it.
It's
easy
to
deploy,
there's
no
scripting
involved.
You
know
all
the
docker
files
builds
all
that
it's
auto-scaling.
Without
you
doing
anything
about
it.
You
know
priced
per
invocation
and
you
don't
need
to
think
about
how
to
do
the
event
triggering
you
know.
Api
gateway
is
baked
in
for
the
cloud
provider
also
other
messaging
and
streaming
solutions
are
baked
in.
C
You
just
need
to
write
your
business
logic
and
it
handles
versioning
lifecycle,
etc.
If
you
compare
it
with
legacy,
serve
solutions
around
the
SOA
architecture,
which
will
serve
as
the
predecessor
is
a
sort
of
a
successor
of
that
there
were
many
design
for
one
language
like
Java,
had
has
a
bunch
of
streaming
solutions.
You
know
storm
and
others.
There
were
also
other
solutions
in
the
past
that
were
very
language
specific
and
usually
served
Java
like
and
that's
another
thing.
The
service
sword
solves
lets.
C
You
work
with
your
choice
of
language,
but
there
are
some
bad
things.
Typically
service
is
extremely
slow.
We're
talking
about
tens
of
milliseconds,
if
not
hundreds
of
milliseconds
per
invocation,
very
low
concurrency.
You
know
throughput
of
maybe
thousand
events
for
a
second
in
most
of
the
implementations
out
there
because
of
the
way
they
were
architected.
In
many
cases
the
you
know
there
are
solutions
that
actually
launch
the
process
every
time,
there's
a
an
event.
I
call
it
CGI
CGI
starless,
but
there
are
there
a
better
solution,
but
still
there's
really
lack
of
concurrency.
C
Better
key
challenge
is,
if
you're,
using
a
cloud
provider
specific
stir
this
framework,
like
lambda
you're,
only
limited
to
data
storage
that
was
certified
by
that
service
function
like
Kinesis
for
streaming.
If
you
happen
to
want
to
use
a
Kafka
or
RabbitMQ
or
something
else,
you
can't
use
it,
you
can
only
use
the
specific
data
and
events
forces
that
are
supported
by
this
provider
and
the
other
open
source
solutions.
You
will
find
that
it's
very
limited
choice
only
like
HTTP,
potentially
some
limited
support
for
Kafka.
C
There
are
also
other
challenges
around
debugging
and
diagnostic
and
build
dependencies,
we're
not
going
to
cover
those.
We
hope
they
will
be
sold
over
time
and
there
is
a
challenge
of
lack
of
standardization
and
common
API
is
when
you
go
to
one
platform,
essentially,
it's
very
different
than
the
other
platform.
Keeping
it
within
the
same
platform
you
may
have
very
different
API
is
across
different
sources
and
solutions.
C
This
is
really
the
role
of
CN
CF
in
that
landscape,
and
this
is
a
new
diagram
that
urban
CN
CF.
We
launched
around
the
service
landscape
to
show
you
all
the
players
and
the
solutions,
and
what
we've
done
is
for
the
white
paper
that
I
helped
co-author
is
defining
the
space
requirements
trying
to
normalize
the
views
across
the
different
providers
and
that
led
to
a
new
standard
that
were
working
on
something
called
cloud
events
which
is
a
way
to
normalize
all
those
events
versus
into
sort
of
a
common
schema.
C
C
So
with
that,
there
are
a
lot
of
things
that
we're
working
on
in
C
and
C
F,
and
if
you
people
want
to
join,
there's
a
there's
called
weekly
calls
and
we
welcome
participation
so
we're
without
you
know,
one
of
the
things
that
we
built
in
order
to
address
those
challenges
that
I
listed
before
around
serverless
from
performance
and
lack
of
conformance
around
different
platforms.
We
created
a
new
stimulus
framework
which
emerged
from
something
proprietary.
We
had
originally
be
quasi
Oda
sister.
C
A
second
generation
that
has
for
a
very
real
time,
function,
processing
engine
design
around,
go
with
some
very
unique
capabilities
about
hooking
to
other
languages,
using
things
like
shared
memory
and
embedding
and
other
like
low-level
tricks,
and
to
maintain
this
performance,
while
you're
still
server
list
so
we're
you
know,
nucleo
can,
and
that
was
their
straight
move.
Multiple
times
can
even
run
like
four
hundred
thousand
events
per
second
on
a
single
process,
that's
more
than
a
hundred
times
faster
than
most
other
solutions
and
that's
in
go.
C
Another
thing
that
we've
done,
which
is
now
going
to
be
standardized
as
part
of
the
cloud
events
efforts,
is
creating
a
normalized
way
of
delivering
events
to
the
function
and
that's
using
something
we
call
you
know
the
event
triggers
are
all
fitting
into
from
a
listener
and
running
the
function.
So
it
doesn't
matter
if
your
event
is
an
HTTP
request
or
a
message
queue
from
you
know:
RabbitMQ
or
a
cloud
provider
streaming
solutions
like
in
easy
storage
or
event
hub
or
a
Google
pops
up.
C
All
those
are
supported,
and
eventually
your
function,
gonna
get
exactly
the
same,
API
call.
So
that
means
that
now
you're
you
have
this
portability
there
and
another
thing
that
we've
embedded
on
is
the
platform
abstraction
layer,
because
when
you're
in
a
cloud
you
want
to
have
the
full
cloud
experience
like
if
you're
using
Azure
instead
of
managing
your
own
kubernetes
by
yourself,
you
may
want
to
actually
leverage
the
AKS
service.
That
summer
is
going
to
speak
about
later,
which
provides
you
manage
a
kubernetes
experience.
Even
some
interesting
pricing.
C
You
know
schemas
where
you
just
start
four
per
container
and
memory
yeah
yeah,
and
so
what
we've
done
is
created
an
abstraction
layer
where,
when
it
sits
on
a
specific
platform
where
we
do,
the
integration
and
over
here
is
working
on
the
native
integration
with
with
ager
and
other
people
from
a
scheme,
including
the
integration
into
the
logging
framework,
may
be
conspicuous
yeah.
So.
B
A
key
feature
in
any
stimulus
platform
is
my
logging
and
monitoring.
So
right
now
we
are
doing
an
effort
to
integrate
the
application
insight,
which
is
a
natural
component
into
nucleo
and
essentially
what
it
gives
you
a
managed
service
to
that
can
get
all
your
logs
and
your
monitoring
data
metrics
and
give
you
insights
on
it.
It
has
some
nice
ml
algorithms
behind
the
scenes
and
yeah,
and
it's
going
to
be
integrated.
B
Natively
in
new
cleo
is
and
but
you
will
be
able
it's
done
in
a
sort
of
way
that
you
will
be
able
to
do
to
integrate
other
platforms
as
well,
so
you're
not
going
to
be
coupled
at
the
old
math.
That's
the
whole
concept
of
new
Pio.
You
are
not
coupled
to
any
service
or
anything
called
provider.
You
will,
you
will
be
able
to
plug
in
whatever
service
you
want.
Yeah.
C
So
if
you're
watching
the
Twitter
feed
on
like
Elsie
at
our
and
other
huge
discussion,
serverless
versus
fast
in
a
function
as
a
service
where
service
is
more
packaged
and
integrated
and
fat
is
more
likely
to
yourself,
one
of
the
thing
that
we
try
to
do
in
in
nuclear's
essentially
provide
both.
So
if
it
is
integrated
into
a
platform
like
a
juror
or
HP
ke,
you
know,
excuse
me
Tom
around
plugging
your
competitors
or
a
juror
or
the
positive
platform
or
anyone
else
that
chooses
to
the
managed
service
around
the
service.
C
C
So
that's
in
general,
there
also
some
other
pretty
cool
features
and
nucleo
very
unique,
like
support
for
screaming
and
batching
for
something
called
dealer.
You
know
data
binding,
which
is
essentially
where
your
your
function
already
gets
a
plug
into
the
data
objects
directly.
So
you
don't
need
to
mess
with
the
Global's
in
functions.
You
know
structured
logging,
versioning,
a
bunch
of
other
nice
features
that
we
thought
a
lot
about.
C
Even
Kelcey
think
it's
also
Brandon
burns
for
me,
not
just
skills.
So
just
again
talking
about
service
and
the
context
of
data
so
serve
a
list
of
many
people.
Use
service
for
HTTP
request
reply
models.
That's
the
most.
You
know,
that's
the
typical
PHP
CGI
nodejs
use
case,
but
people
serve
tend
to
forget
that
you
could
do
a
lot
of
other
interesting
things
with
service.
Not
all
the
service
framework
screens
support
those
models,
mainly
the
club
provider,
ones
or
more
feature.
C
Aids
especially
comes
on
a
major
thing:
Google
is
sort
of
less
functional
and
what
you
have
is
essentially
three
main
invocation
models
in
service.
The
synchronous,
like
the
HTTP
request,
reply
model
where
you
invoke
a
function
and
then
you
get
a
response,
you
don't
really
know
which
function
and
answer
that
request
and
you're
just
getting
the
response
and
another
one
and
one.
C
Or
rabbitmq,
if
you
remember
the
open
source
party,
that's
essentially
like
an
exchange
like
an
email
exchange,
you
throwing
a
message
and
every
function
will
pick
up
one.
It
is
server.
It
delivers,
there's
no
ordering
in
those
messages
they
can
essentially
be
processed
in
any
order.
That's
why
it's
called
sir
an
asynchronous
structure.
It's
you
and
then
there
is
another
model
which
is
streaming
model.
It's
different
than
a
synchronous
model,
most
of
the
open
source
frameworks.
C
Today
they
support
only
HTTP
and
the
some
support
may
use
Kafka,
but
they
essentially
use
Kafka
in
the
asynchronous
mode.
With
a
single
shark
and
a
bunch
of
functions
they
read
from
the
single
shark.
So
it's
not
really
high
performance.
What
do
you
want
to
do
for
streaming
and
that's
something
that
a
Adria
and
Amazon
and
a
nucleo
support?
Is
the
ability
to
have
multiple
assured,
multiple
partitions
of
the
stream
and
those
partitions
will
be
consumed
by
multiple
function?
Instances?
So
that's
providing
you
with
parallelism.
C
There
is
a
big
challenge
in
streaming
because
everything
has
to
have
been
in
order
and
there's
a
notion
of
checkpointing.
So
something
needs
to
manage
it
and
in
the
other,
open
source
project.
There
is
nothing
like
that
and
ok
there's
something
called
dealer:
it's
still
not
in
the
main
GA
trunk
that
we
hope
to
bring
it
there
very
soon,
but
it
is
a
pretty
cool
feature
that
allows
you
to
dynamically
checkpoint
and
move
partitions
and
other
batch
of
batch
tasks.
C
C
The
in
lambda
I've
seen
a
lot
of
customers
struggling
with
today
to
take
one
job
and
paralyze
that
across
multiple
processes
and
give
each
of
the
processor
a
piece
of
the
job,
and
so
that
tell
us
about
the
invocation
mode
when
you
were
talking
about
service
needs
to
understand
that
there
is
serve
an
automated
deployment
model
in
some
cases,
it's
a
single
step.
In
some
cases
it's
multiple
steps,
so
it
all
starts
with
a
code
and
some
description
of
what
you're
trying
to
do.
C
This
description
could
be
a
mo,
it's
HTTP
request
or
anything
else
and
those
things
the
spec
and
could
essentially
instruct
the
Builder
how
to
build
an
artifact.
In
most
cases,
the
artifact
of
the
function
is
stored
as
a
docker
container
locally
or
remotely,
and
then
this
docker
container
gets
deployed
onto
a
platform
onto
a
cluster
and
once
it's
deployed
on
the
cluster,
there
are
things
that
make
sure
that
it's
elastically
scales
in
and
out
gets
events
gets.
C
The
events
are
being
load
balanced,
all
the
stuff,
in
our
case,
also
getting
data
binding
to
the
functions,
etc.
Now,
in
most
solutions,
this
is
a
single
hop
sort
of
a
single
click
operation.
In
some
cases,
you
actually
want
it
to
be
broken
to
two
different
steps.
One
is
that
building
the
function
generating
the
artifact
and
then
potentially
deploying
the
same
artifact
on
ten
different
location,
especially
when
you
have
a
cluster
of
multiple
locations
in
the
edge,
and
you
want
to
take
the
same
thing
and
run
it
on
different
locations.
C
C
C
Another
interesting
thing
is
that
a
new
Clio
has
multiple
platforms
support,
but
with
kubernetes
or
the
kubernetes
very
variants
it
has
native
integration.
So
it's
what
what
is
referred
to
as
CRD
customer
resource
definition,
I,
actually
wrote
the
blog
post
on
teaching
how
to
build
your
own
CRD
for
the
new
stack
and
I
can
sort
of
send
a
link
later
and
on
that.
C
But
what
you
could
see
that
if
I
go
to
a
Carroll,
which
is
the
kubernetes
CLI
I,
can
list
the
functions
and
will
actually
see
all
the
content
and
it
start
preserving
some
kubernetes
resource
semantics,
like
secrets
and
sort
of
resource
and
and
in
addition
adds
other
thing
and
it
has
labels
and
searching
capabilities
based
on
labels,
and
you
know
that.
But
in
addition
it
has
other
unique
features.
C
Like
data
bindings
event
sources,
the
full
spec
would
be
found
on
sort
of
the
negita
aboard
the
website
and
also
another
unique
thing
is
that
you
can
define
the
number
of
replicas
in
most
places.
You
have
either
like
start
from
zero
or
it's
fully
dynamic,
but
we
hear
more
and
more
people
saying
you
know
what
I
may
want
to
have
a
variable
size
or
well-defined
umber
of
functions,
running
or
a
scope,
a
minimum
and
a
maximum
and
not
go
below
or
above
that.
So
that's
another
thing.
We
we've
taken
care
of
yeah.
C
So
that's
that's
from
that
and
let
me
talk
about
some
use
cases
that
relates
to
the
data
science
and
so
one
example
that
I'm
going
to
also
show
you
a
short
video
in
a
minute
is
a
scenario
where
I
have
data
being
ingested
by
you
know:
Thomas
car
connected
Cora's,
ride-sharing
service.
We
have
a
bunch
of
those
type
of
customers
and
and
this
data
needs
to
be
served,
aggregated
potentially
contextualized
with
other,
for
example,
location
data
from
a
Google
map
service.
C
In
this
case,
so
we
essentially
run
a
function
that
intercepts
the
calls
from
cars
and
and
devices
contextualize.
It
has
stored
the
data
in
a
server.
No,
you
know
no
sequel
I
do
there.
The
next
thing
that
you
want
to
do
is
take
all
that
data
that
was
ingested
and
run
our
machine
learning
prediction.
C
On
top
of
that
data
to
potentially
see
future
trends
and
once
you've
done
that,
you
may
want
to
do
other
queries
on
that,
but
the
cool
thing
that
we
did
in
this
case
instead
of
running
sequel
queries,
which
is
pretty
boring
today
we
are
talking
sequel
or
we're
talking
to
the
system,
and
the
system
will
automatically
respond
with
voice
responses,
as
well
as
graphical
responses
of
the
results.
So,
essentially
you
created
something
which,
up
to
a
couple
of
years
ago,
was
totally
science
fiction.
You
created.
C
C
A
A
C
So
you
see
this
was
something
that
I
think
up
to
like
some
time
ago
was
impossible.
I
think
many
people
that
deal
with
data
still
think
sequel,
and
this
is
just
opening
the
you
know
the
imagination
to
show
you
that
you
can
start
using
voice
and
text
and
video
and
and
chat
box
in
your
system.
But
this
is
really
where
serverless
shines,
because
they
guys
now
no
longer,
you
know.
Urls
and
HTTP.
C
They
need
to
decide
if
there's
going
to
be
a
failure
and
to
do
some
corrective
measurement.
So
one
of
the
things
that
we've
we've
done
is
essentially
taking
all
the
sensor
data
that
flow
in
real
time,
including
alerts
and
telemetry
and
matrix,
and
all
that
feeding
it
into
a
function
using
the
streaming
protocols.
C
Then
this
function,
essentially
what
it
does
is:
building
buckets
and
aggregates
sexually
collecting.
You
know
some
and
standard
deviation
and
other
information
into
arrays.
We
didn't
serve
our
no
sequel
and
and
that,
if
it
finds
some
immediate
anomaly,
doesn't
need
to
be
too
clever
and
run,
a
I
could
immediately
alert
another
system
the
same.
There
could
also
be
visualized
using
some,
no
sequel
API
in
a
desk
work.
The
next
step
is
again
to
take
this
periodically.
C
Those
vectors
that
accumulate
all
sorts
of
statistical
information
like
some
and
variants
and
averages
and
feedback
together
with
other
information
like
the
machine
learning
model
machine
made
the
data.
When
was
it
serviced
the
last
time
environmental
data
like
the
weather
in
this
specific
location,
etc?
All
that
is
something
that,
for
is
what
is
called
referred
to
in
AI,
a
feature
vector
or
an
enriched
feature
vector
which
collects
the
event
related
details
and
all
the
other
environmental
and
contextual
data
that
is
used
in
order
to
take
the
prediction
decision
now.
C
In
order
to
run
the
prediction,
you
need
the
machine
learning
model,
it's
very
efficient
to
build
a
machine
learning
model
in
the
cloud,
not
necessarily
in
the
factory,
because
in
the
cloud
you
have
lots
of
machines
and
storage
and
you
could
take
five
year
worth
of
data
and
analyzed
it
and
that
build
a
model.
So
what
we've
done
here
is
is
we
built
the
model
in
Asia
using
something
called
HTML
workbench,
and
this
model
was
actually
shipped
as
a
container
into
the
same
kubernetes
cluster?
C
Is
this
machine
going
to
fail
or
or
when
it's
error
and
if
it
is
going
to
fail
that
immediately
generates
an
alert
so
I
don't
have
to
go
and
send
all
that
data
to
the
cloud
and
back
in
order
to
take
a
decision?
I
can
take
a
real-time
decision
based
on
that
data.
Now
the
thing
is
that
we
do
need
to
interact
with
the
cloud.
C
So
we
have
two
functions
here:
one
of
the
functions
that
essentially
takes
the
matrix
that
are
being
aggregated,
not
the
road
data,
but
essentially
the
compressed
and
aggregated
implementation
to
save
bandwidth
and
all
that
and
every
few
hours
it's
going
and
sending
that
into
a
blog
store
in
the
cloud.
Okay
and
and
then
the
the
cloud
can
go
and
do
the
reinforce,
learning
learn
some
information,
because
once
you've
done,
you
know
you
can
once
you
have
real-time
results.
If
you
wait
a
bit,
you
see
if
those
results
really
happen,
you're
predicting
a
failure.
C
You
wait
an
hour
and
you
see
if
there
was
a
failure,
so
you
can
learn
from
that
information.
That's
called
reinforce
learning
so
that
data
can
essentially
go
and
update
the
machine.
Learning
model,
so
the
cloud
can
assist
you
just
put
the
blob
in
a
blob
store
in
an
object
store,
and
this
put
in
the
in
the
black
fur.
Over
object,
store
will
generate
a
web
trigger
web
book
and
that
web
book
can
essentially
trigger
a
function
that
takes
that
machine
learning
model,
as
for
is
it
back
into
the
platform
and
reinitialize
the
model.
C
So
what
you
see
here,
I
didn't
use
anything
else,
beside
service
functions
and
a
no
sequel
database
to
build
a
complete
solution
and-
and
that
thing
interacts
with
with
the
cloud
services
in
a
very
clean
way,
no
need
to
manage
virtual
machines,
no
need
to
install
your
Oracle
of
choice,
your
my
sequel,
you
know
deal
with
hard
drives
and
SSH.
None
of
that.
C
B
B
C
Another
is
so
we
see,
we've
talked
a
lot
about
use
cases
of
more
traditional
data.
Another
use
case
that
we
have
is
actually
processing
images
and
videos.
You
know
think
about
things
like
the
Boston
bombing.
You
know
it
took
two
weeks
after
the
event
to
figure
out
some
information
and
and
if
you
essentially
can
feed
all
the
information
from
the
video
cameras
and
analyze
it
in
real
time,
you
may
actually
be
able
to
predict
or
or
stop
an
event
before
it
even
occurred.
C
So
those
thing
required
first
thing
to
essentially
take
those
images
and
ingest
them.
When
you
ingest
the
pictures,
you
may
extract
metadata
information.
Like
you
know,
the
JPEG
Hagar's
geolocation
of
the
picture
is
other
metadata
information,
color
schemes,
etcetera
once
those
are
stored,
is
for
like
an
object,
store
or
other
data
storage
systems
that
can
trigger
an
event
to
other
perceptive
functions
that
embed
something
like
tensorflow
and
using
deep
learning.
C
They
can
analyze
those
pictures
and
try
and
extract
other
meaningful
information
from
the
actual
pictures
using
a
pre
learn
the
model
once
you
have
that,
and
you
generate
this
metadata
and
you
write
it
back
to
the
database.
This
metadata
can
be
queried
by
analytic
tools
to
start
looking,
for
you
know,
maybe
we've
seen
the
same
guy
in
the
same
location
five
times
and
last
week.
C
Maybe
we've
seen
a
guy
moving
across
five
different
cameras
and
having
sort
of
trends,
so
those
things
can
be
analyzed
using
other
analytic
tools
and
eventually
the
system
actually
generate
triggers
that
will
send
an
SMS
or
an
alert
to
someone
to
deal
with
that
same
for
you
are,
you
can
have
functions
or
the
sense
you
just
bring
the
UI
and
again
the
eye
could
be
text
could
be
video
audio.
You
know
anything
chat
BOTS.
C
So
this
is
another
thing
where
you
see
that
server
list,
which
is
stateless
and
very
elastic
couple
with
some
managed
data
services
to
provide
you
a
very
powerful
and
rapidly
deployed
solution
so
well
without
a
lamb.
I'll
show
you
some
some
demos
of
nuclear
in
the
context
of
okay.I
and
I've.
Also
written
some
blog
post
on
the
in
hacker,
no,
no
with
actual
code,
examples
that
if
anyone
haven't
seen
and
nucleo
site-
and
you
have
all
the
information
there,
there's
also
no
Kyoukai
outside-
where
you'll
see
more
information.
C
Now,
when
you
open
new
Cleo
and
keeps
on
a
process
thing,
you
know
advancing
on
very
rapid
database.
We
have
sir
very
strong
developers,
including
help
of
summer
and
Iran
and
Mickey,
and
a
bunch
of
a
very
strong
members
and
we're
now
adding
more
languages,
improving
the
UI
quite
significantly
in
the
next
coming
versions.
What
you
see
here
is
the
new
Clio
UI,
which
is
still
in
us
or
what
we
define
as
playground,
but
in
very
soon
it's
going
to
be
a
full-fledged
UI
with
also
monitoring
and
hierarchies
and
pretty
powerful.
C
What
you
can
see
in
a
nuclear
UI
that
when
you
open
it
and
you
can
install
new
Cleo
just
using
a
single
docker
command,
you
don't
really
need
kubernetes.
You
can.
If
you
go
to
the
website,
you
have
an
instruction.
It's
essentially
just
typing
this
command
in
your
Mac
or
your
Linux,
and
you
have
a
nuclear
running.
Just
open
480
17,
you
have
also
even
on
Windows
yeah.
A
C
C
C
For
example,
if
you're
going
and
configuring
HTTP,
you
can
go
and
put
all
your
API
gateway
definitions
already
here
you
don't
need
to
go
on
a
separate
tool
and
and
if
you're
going
to
to
open
the
kubernetes
cluster,
that
is
backing
this
system
up.
You
will
actually
see
that
you
have
nuclear
functions
and
services
and
deployments
and
ingress
rules
in
the
case
of
kubernetes
for
the
api
gateways,
it's
automatic
and
it's
all
done
for
you,
you
don't
need
to
think
about
all
of
that.
So
that's
this
thing.
C
Other
other
examples
that
you
have
is
like
tensorflow,
which
is
for
image
recognition,
and
there
is
another
nice
example
of
face
recognition
using
the
azure
face
recognition
API.
It
does
require
some
face
api,
some
keys
that
you
need
to
for
free
for
microsoft
it.
Actually,
there
is
a
full
description
on
how
to
obtain
those,
but
this
is
like
to
do
face
recognition.
So
with
that,
I
won't
take
too
much
time
out
of
a
tourist
demo
and
yep.
B
C
B
So
essentially,
Microsoft
managed
the
Masters
in
kubernetes
and
you
don't
pay
for
them.
We
guarantee
SLA
for
that.
You
can
scale
it
quite
easily
with
a
single
command
and
it's
also
integrated
with
the
other
ecosystem,
extensions
that
we
have
to
provision
other
resources,
and
let
me
show
you
the
demo
all
right
so
one
second.
B
B
So
you
can
find
it
only
in
the
nuclear
demos,
repo
and
essentially
what
we're
doing
here.
We
want
to
simulate
a
lot
of
cars.
Okay,
we
want
to
get
determine
the
metrics
through
them
and
specific
to
this
scenario.
We
want
to
get
the
tire
pressure
now,
with
this
tire
pressure.
We
want
to
alarm
the
system
if
we,
if
some
of
the
drivers
have
problems
with
the
tire
pressure,
but
tire
pressure
can
change
according
to
weather.
B
So
let
me
just
go
here
so
the
first,
the
first
function
that
we
have
here:
okay,
that's
a
fleet
sim
and
essentially
it's
a
it's
a
it's
a
simulator.
It
is
it's
a
simulator
that
sends
a
lot
of
metrics
through
as
your
event
hubs
to
the
to
the
functions
in
nucleo.
So
you
can
see
here.
The
function
is
everything
go
but
right
here
at
the
top
from
line
number
17
to
37,
you
can
see
the
definitions
of
the
function.
B
C
Want
to
make
one
point
here
is
that,
in
order
to
make
life
simpler,
we
allows
instead
of
having
a
separate
Yama
file
and
uija
and
CLI
commands,
and
all
that
you
can
essentially
go
into
the
function
and
put
it
the
creation,
as
a
defines
all
is
so.
That
means,
if
you
email
this
function
to
Sangha
and
runs
it
and
it's
all
nuclear,
it
will
actually
have
the
full
configuration,
but.
B
You
can
have
anything
in
different
EMF.
Also,
you
can
see,
we
converging
it
or
something
like
that.
So,
but
I
won't
go
to
through
the
code,
but
we
generate
a
lot
of
data
and
we
send
it
to
a
jury
when
hub.
That's
one
function
now
the
second
function
that
we
looked
that
we're
looking
at
it's
the
sensor
handler
this
function
gets
it
receives
the
stream
from
from
a
driven
hub
process,
the
data
it's
send
and
it
sends
it
sends
it
to
another
function,
called
weather.
B
Okay-
and
here
you
based
on
the
location
of
the
car,
we
are
generating
some
random
weather
condition,
okay
and
send
it
sending
it
back
to
this.
Is
the
sensor
handler
function,
okay
and
based
on
that,
we
will
have.
Let
me
go
right
here,
based
on
that
we
will
have
different
alarms.
Okay,
all
right,
so
let
me
show
you
the
actually
demo,
so
I
recorded
the
function,
so
it
will
first
start
by
showing
you
the
cue
control,
get
note
which
this
is
my
a
KS
cluster.
Okay,
so
you
can.
B
C
B
Tooks
me
literally
like
two
minutes,
and
you
can
see
that
everything
is
that
I
have
two
parts.
Okay,
and
now
we
will
install
the
functions
okay,
so
the
first
command
will
be
a
new
cuttle
deploy
weather
and
the
weather
file.
The
weather
go
file,
okay
and
it's
building
in
the
image.
It
takes
a
couple
of
seconds.
Okay
and
then
it
will
push
it
into
Azure
container
registry,
which
is
my
private
container
registry,
because
I
don't
want
my
functions
to
be
up
grabbed
in
the
public.
C
B
So,
in
a
majority
of
cases
you
can
use
a
third
parties
like
from
aqua
and
other
services,
to
scan
your
images
and,
of
course
it
everything
is
backed
by
the
whole
as
your
security
mechanism,
and
so
this
is
the
second
function
that
you
see
now
you
see
that
it's
deployed.
Okay
and
you
will
see
the
third
function
in
a
second
okay
and
then
we
will
log
into
it.
So
this
is
the
flight
same.
C
B
B
B
C
Logging
is
non-intrusive,
so
it
has
level
base.
So
you
can
say
right
now:
I
want
to
run
my
function
in
debug
level
or
you
know
ever
level.
So
in
real
time
you
can
change
that
where,
in
places
like
lambda,
if
you
it
records,
even
your
debug
messages,
so
the
debug
message
will
actually
slow
you
down.
Yeah.
B
B
So
now
I
know
the
function
is
working.
Everything
is
connecting
to
it.
I
want
to
go
back
a
little
bit
to
the
function
itself,
so
you
mentioned
before
that
we
don't
want
to
couple
a
function
to
a
specific
service.
Okay.
So
if
you
look
at
the
signature
of
the
function
and
that's
very
interesting,
you
can
see
that
I
have
a
context,
okay,
which
basically
does
the
lot
right.
B
Now
we
have
there,
the
nucleo,
logger,
okay,
which
is
it
doesn't
matter
which
language
you
write
your
function
in,
you
will
have
the
same
interface
and
then
another
thing
that
we
have
here:
it's
a
the
nucleo
event
and
basically
the
nuclear
event.
It's
that
that's
the
event
that
have
all
the
details,
all
the
all
the
properties
from
the
I'm
getting.
C
B
Essentially,
what
this
does
the
the
event
it
doesn't
matter
if
I
connect
as
your
event
hub
or
an
HTTP
request.
Okay,
for
that
matter,
I
will
get
the
same
event,
maybe
a
little
bit
of
different
data
inside.
But
the
interface
will
be
the
same,
so
you
can
actually
test
this
function.
This
function
with
an
HTTP
always
a
your
event
hub
now
also
another
stuff
that
I
have
here
with
the
data
bindings.
B
So
after
everything
is
set,
okay,
so
I'm
sending
everything
to
the
enriched
enriched
fleet
event
hub
okay,
which
is
after
we
correlate
all
the
weather
and
the
tire
pressure
conditions
I'm
sending
in
the
and
sending
that
as
it
is
back
to
event
hub,
because
why?
Because
I
want
to
send
it
back
to
spark
okay
and
then
spark
will
receive
it's
going.
It's
going
to
event
hub,
okay
and
then
I
want
to
process
the
messages
there.
So
it's
fine
you're,
also
demo
scenario
that
you
can
see
here,
I'm
connecting
to
event
hubs.
C
B
So,
actually
that's
to
connect
to
the
nuclear
playground
in
Azure.
It's
quite
easy.
Basically,
just
a
very
simple
port
forwarding,
okay
and
then
you
can
have
local
office
because
you
don't
actually
want
it
exposed
with
a
load
balancer
and
public
IP.
So
you
want
to
make
it
more
secure.
So
if
you
have
access
to
cuba
to
keep
control,
you
will
have
access
to
nuclear,
but
that's
like
protecting
your
SSH
keys.
B
C
B
And
on
the
CLI,
basically
very
simple,
so
the
CLI
is
that
it's
reading
go
very
simple
and
then
just
commands
to
do
a
deploy.
You
can
see
here
that
you
just
specify
the
name
of
the
function:
okay,
the
the
path
of
the
go
file;
okay,
the
namespaces
that
you
want
in
kubernetes
for
the
for
the
for
the
for
the
function
to
be
deployed,
and
then
the
runtime
of
the
function
now.
C
B
C
That
means
that
you
can
create.
You
know
open
your
editor
write.
Some
functions
commit
to
you,
get
up
and
run
the
deploy
process
again
we're
working
on
enhancing
that
even
even
further.
So
there
is
also
another
testing
framework
that
I
actually
built
that
allows
you
to
run
the
within
the
editor
itself
within
IntelliJ
or
vs
code
with
you
know,
with
breakpoints,
with
everything
that
emulates
the
wrapper.
That's
something
we
call
you'll
find
it
in
the
repo
I
think
it's
something
I
think.
A
C
C
C
A
C
The
director
of
different
solutions,
you
know
the
openness
comes
with
with
IBM
open
risk
is
not
as
performance
as
nucleo.
The
design
is
quite
different.
If
you
know
how
it's
working,
it
has
essentially
serve
an
asynchronous
design
which
gets
a
higher
latency
and
slower
performance.
Also
in
terms
of
deployment
open,
it
quick
choirs,
a
lot
of
dependencies,
you
know,
go,
try
installing
it
and
what
you've
seen
the
Tomer
is
essentially
installed
nuclear
in
two
minutes,
because
it
leverages
like
it
CD
on
kubernetes
cluster
for
persistency.
The
kubernetes
CRD
is
the
suit
kubernetes
api
services.
C
So,
essentially,
that's
why
it's
so
easy
to
install
because
it
leverages
the
platform
a
specific
thing.
So
it's
not
complementing
openly
it
serve.
Another
solution
that
you
need
to
choose
between
the
two
nuclear
has
a
really
a
broader
feature
set
when
it
comes
to
data
analytics
and
not
so
much
higher
performance,
so
obviously
we're
biased
towards
that
yeah.
C
So,
for
even
embedding
so,
for
example,
in
lukio,
the
Python
runtime
actually
resides
as
part
of
the
go
process.
There's
not
even
an
rfp,
see
between
the
go
wrapper
that
implements
the
events
and
actual
Python
function,
which
means
that
the
Python
functions
are
running.
A
nucleo
wrapper
will
run
faster
than
you're
running
your
own
Python
functions.
We
managed
to
achieve
about
80,000
item
either
HTTP
requests
per
second
being
processed
on
Python,
which
was
about
four
times
faster,
nor
what
we
managed
to
do
with
Python
without
service.
C
A
C
Trial,
so
there
are
a
bunch
of
companies
using
it
in
the
server
sonars
that
are
not
sort
of
like
mission
critical
right
now.
Now
we
have
committed
to
to
them
that
around
April
timeframe,
April
May
time
frame,
there'll
be
a
production
were
degrees.
We've
invested
quite
a
bit
on
production
related
features.
Recently,
it'll
you
see
where
added
like
liveliness,
who
can
readiness
hooks
and
right
now.
You
know,
for
example,
we
validate
that
when
the
run
time
is
up
the
function
is
everything
is
reporting
that
it's
up
and
only
them
which
were
responding.
C
There
are
a
lot
of
issues
and
things
that
we're
dealing
with
around
AJ
and
robustness,
and
and
also
ramping
up
the
UI
to
something
which
a
full-fledged
air
management
with
monitoring
support,
and
you
know
being
able
to
carve
projects
namespaces
all
that
stuff
and
security.
So
we
are
working
towards
around
April
May
time
for
a
full-blown
solution,
which
will
be
much
more
feature-rich
than
the
other
solutions
in
the
market,
but
with
our
already
people
that
use
it
in
production
right
now,
just
in
limited
scenarios.
A
Well,
thank
you
so
much
for
your
time
today,
gentlemen,
greatly
appreciated
we're
gonna
turn
this
video
round
about
a
day,
so
look
for
an
email
with
the
link
to
YouTube
will
also
share
out
some
of
the
information
balled
out
as
far
as
like
the
big
is
repos
and
contact
information
and
all
that
stuff
tome,
Bern
general,
your
own.
Thank
you
so
much
for
your
time
today.
Thank
you
to
all
the
attendees
for
participating
and
we'll
see
you
soon
again
on
the
next
series
of
CNC,
app
webinars
and.