►
Description
Speakers: Tushar Katarki, Itay Gabbay
AI is popular and yet faces two big challenges in the industry: 1) self-service and automation 2) Use in real production.
At the Israel Ministry of Defense we are taking on the challenges with containers and Kubernetes. We have built AI-as-a-service with open source tools and Kuberentes. Our Data Scientists use the service for data, experimentation and to deliver models into production iteratively with self-service and automation.
Using Kubernetes, we are able to run massive machine learning pipelines automatically, and improve our machine learning models. We implemented several principles of AutoML - a wide research area nowadays. Using AutoML & Kubernetes, we can further improve our machine learning models and pipelines - automatically.
Come find out how we built our AI service on Kubernetes, issues we ran into and best practices with a live demo and supporting slides.
A
So
our
next
session
is
gonna,
be
starting
right
now
it's
gonna
be
a
case
study
given
by
two
shark
authority
who
work
down
AI
as
a
service
platform
on
open
shift
using
out
ml
and
equipment
at
ease.
They
are
able
to
run
massive
machine
learning
pipelines
and
improve
the
machine
learning
model
automatically
extremely
interesting
topic.
B
Thank
You
Chun
and
thanks
everybody
for
coming
here.
Welcome.
I'm
super
excited
to
be
here
and
I'm,
going
to
start
my
stop
clock
here
so
that
I
keep
a
track
of
time.
I
have
good
amount
of
material
to
cover
hope
you
are
all
having
a
good
time
here
in
San
Diego,
in
spite
of
the
rain
and
the
one
or
two
days
we
are
here,
it
decides
to
rain.
So
that's
great,
but
this
is
going
to
be
a
much
more
interesting
topic.
So
it's
good
that
you
are
here.
B
So
let
me
introduce
myself
and
my
let's
start
with
ETA
ETA
gabay
is
the
machine
learning
lead
for
a
team
of
team
in
the
Ministry
of
Defense
for
Israel
and
an
ETA
I
have
worked
on
this
presentation.
This
is
obviously
based
on
the
work
that
he
and
his
team
have
done
at
the
mo
D.
Unfortunately,
he
couldn't
make
it
today,
so
I
am
going
to
be
presenting
this
content
on
his
behalf.
Who
am
I.
I
am
too
shark
a
turkey
I'm
a
product
manager
for
open
shift
at
Red.
B
Hat
I
have
been
there
at
Red
Hat
for
several
years
now,
and
one
of
the
things
that
I
like
about
my
job
is
that
I
get
to
talk
to
a
lot
of
customers.
Customers
who
are
really
using
our
product,
kubernetes
openshift,
obviously
is
built
upon
that
and
that's
very
exciting.
You
know
I
see
a
lot
of
different
use
cases,
the
problems
that
they
run
into
the
escalations
that
they
bring.
It's
all
a
good
challenge,
so
one
of
the
favorite
customers
of
ours
and
mine
particularly,
is
the
Ministry
of
Defense.
You
know.
B
Obviously
this
is
a
large
governmental
organization
and
what
they
are
what's
happening.
There
really
is
that
leaves
the
click
owner.
They
are
actually
doing.
They
are
ongoing
digital
transformation
and
a
lot
of
companies
are
doing
that
they're
doing
it
with
private
cloud
in
mind
and
then
one
of
the
major
things
that
they
want
to
do
really
there
with
the
additional
transformation,
is
accelerate
application,
delivery,
accelerate
R&D
projects
and
particularly
of
interest.
Obviously,
for
this
talk
is
what
they
have
done
to
do:
accelerate
machine
learning
with
this
machine
learning
as
a
service.
B
In
this
particular
case,
it
has
transformed
that
data
into
a
three
dimensional
space
in
which
you
can
put
this
data
plane
in
which
access
the
dish
in
surface
and
therefore
you
are
now
able
to
classify
right.
So
this
is
a
simple
example
to
illustrate
the
point:
if
you
take
it
the
next
step
forward
deep
learning
now,
you
know
actually
I
remember
doing
in
grad
school
undergraduate,
school,
doing
computer
vision
and
object
detection.
Several
years
ago,
20
years
ago,
I
am
reading
myself,
but
the
object
detection.
B
Since
then
not
that
kind
of
traditional
object
detection
was
more
about.
Oh,
is
this
a
yellow
bus
at
the
most
right,
whereas
you
know
in
the
deep
learning
context
what
you
are
trying
to
give
it
context
like
as
humans?
You
know
you
when
you
look
at
this
picture
on
the
left:
you're,
not
just
thinking
about
the
yellow
bus,
but
you're
also
thing
about
the
road
and
the
trees
there
and
the
grass
etc
on
the
right
side.
B
For
example,
it's
a
living
room,
you're
thinking
about
is
it
well
organized
is
eclectic
Exeter
Exeter,
so
humans
actually
think
more
complex
and
that's
something
that
deep
learning
can
do.
Deep
learning
can
kind
of
think
like
humans
and
an
extreme
example
of
that.
Obviously-
and
you
know,
this
is
when
the
the
google
deepmind
in
the
challenge
match
defeated
the
good
champion
a
few
years
ago
a
couple
of
years
ago,
so
this
just
is
at
least
an
illustration
of
how
far
we
have
come
and
what
the
possibilities
are
with
AI
machine
learning
and
deep
learning.
B
So
that's
why
this
is
exciting.
This
is
not
just
a
science
project
anymore
and
therefore
we're
so
now,
let's
look
at
what
some
of
the
characteristics
of
machine
learning
what
is
so
special
about
them,
and
that
will
build
this
theory
towards
what
then
emery
had
to
do
as
a
result
right.
The
first
one
really
is
and
I
think
we
recognize
this
most
of
us
is
data.
They
are
data
and
compute
intensive.
So
how
are
the
data
intense
you?
The
traditional
methods
here
shown
by
the
red
are
on
the
bottom
yeah.
B
Of
course
it
needs
some
data,
but
then,
as
you
see
there
as
the
XP
as
you
throw
more
data,
the
accuracy
doesn't
increase
any
any
much.
But
if
you
look
at
the
top
graph
there,
the
top
line,
the
large
neural
net,
you
know
using
deep
learning
the
accuracy
of
the
model
increases
exponentially
as
the
number
of
data
you
throw
at
it
right.
So
this
is
why
you
know
these
deep
learning
models
are
what
are
known
as
power-hungry
right.
B
So,
similarly,
on
the
compute
side,
just
going
back
from
you
know
this
only
charts
from
2013
to
2018
look
at
the
amount
of
computational
capacity
that
these
various
machine,
learning,
algorithms
and
setups
have
I
have
done
here.
There's
a
300,000
times
increase,
as
you
can
see
on
this
slide,
so
net
net.
We
know
that
they
have
data
and
compute
intensive.
What
is
they
are
very
iterative.
You
know
this
is
a
familiar
model
or
a
pipeline
for
work
flow
for
most
machine
learning,
familiar
with
it.
This
preparation
of
data
there
is
feature
extraction.
B
That
is
you
train
models.
You
are
late
models,
so
what
that,
why
do
you
need
to
do
it
in
a
nitrate?
Your
fashion
is
because
this
just
as
an
illustration,
this
data
here
on
the
left
side
is
just
it's
just
a
tabular
data,
but
it
is,
it
can
be
dirty.
It
can
have
duplicates
you
know.
So
you
know
you
need
to
kind
of
I,
trait
oblique,
lean
it
right
and
and
filter
it,
and
and
and
and
split
the
data
into
kind
of
what
you
need
for
model
building
and
then
for
a
testing
right.
B
B
It
have
a
parameter
tuning,
and
then
you
have
constraint,
cost
constraints
that
you
need
to
manage,
and
then
this
one
just
shows
in
a
gradient
desert
descent,
for
example,
there
could
be
different
lows
right:
there
could
be
local
optimization
point,
so
you
mean
that's,
not
the
one
you
might
want
be
looking.
So
if
there's
a
lot
of
iteration,
as
you
can
see
in
this
lot
of
different
data
is
also
used,
then
then,
finally,
you
deployed
your
model
into
your
satisfied.
B
It
meets
your
criteria,
you
have
deployed
it
into
production,
but
then
also
you
can
see
here,
there's
an
init
drift
right.
So
this
is
an
illustration
of
how
that
drift
is
happening.
The
data
that
is
coming
in
simple,
yellow
data,
initially
has
gradually
changing
changing
over
to
blue
data,
so
you've
got
to
make
sure
that
your
model
is
able
to
account
for
it
right.
So
this
is
drift,
and
so
you
need
to
be
able
to
monitor.
So
all
in
all
ml
is
iterative
right.
B
The
other
thing
that
needs
to
happen
is
machine
learning,
most
people
think
about
data
science
or
data
scientist,
doing
machine
learning
modeling,
but
there's
more
than
that
to
that
right.
First
of
all,
obviously
you
know
it
is
data
scientists
who
want
to
collaborate
with
each
other.
They
want
to
exchange
models
easily.
Then
there
are
other
people
involved
in
this
right.
There
could
be
data
engineers
who,
who
provide
the
data
pipelines.
B
There
could
be
business
analysts
who
decide
or
who
at
least
have
input,
is
to
what
business
problem
you're
trying
to
solve
and
what
you
are
trying
to
optimize,
etc,
and
then
there
are
developers.
You
know,
developers
need
to
take
all
this
and
make
them
available
as
an
application
is
part
of
an
overall
application.
B
So
so
it
is
a
very
collab
and
then
you
want
to
be
able
to
share
it
and
just
bring
it
into
production
right.
So
this
is
a
very
collaborative
and
sharing
model
that
characteristics
that
we
have
it
with
there.
So
so
what
was
the
journey
with
Emily,
so
the
top
three
requirements
based
on
the
tee
up
that
I
did.
You
can
imagine
this
shouldn't
come
as
a
surprise.
They
wanted
access
to
Elastic
Compute.
They
needed
that
compute
resources
they
need
access
to.
They
want.
They
want
the
ability
for
models
to
go
into
production.
B
They
want
to
optimize
and
scale
this
right
so
and
we'll
talk
about
autumn
ml
later
in
that
context.
So
if
you
kind
of
put
it
in
a
nutshell,
what
what
does
the
data
scientist
want?
Right,
like
a
data
scientist,
wants
a
cloud
like
experience.
You
know
where
they
have
access
to
all
the
various
modeling
frameworks.
You
know
that
you
know
this
is
a
very
fast
emerging.
B
You
know
field
and
therefore
new
models
and
new
modeling
frameworks
are
coming
out
all
the
time,
so
they
want
the
latest
and
greatest
they
want
access
to
data.
They
want
access
to
compute,
and
this
includes
GPUs
etc,
which
was
the
case
in
in
the
case
of
M
Modi,
and
then
they
want
to
be
able
to
share
and
collaborate
and
then
bring
these
things
into
production
or
whatever
production
means
in
that
particular
organizational
setting,
so
that
they
can
actually
deliver
value
right.
B
So
there
is
a
stat
out
there
somewhere
that
you
know
lot
of
these
models,
don't
go
into
production,
so
that's
something
that
they
didn't
want
not
to
do
right,
so
they
wanted
to
put
this
into
production,
and
so
this
is
a
conceptual
model
in
some
ways
of
what
they
wanted
to
get
done
right.
So
this
is
really
a
self-service
portal
use
here
in
the
middle,
and
then
you
did
a
scientists
or
a
set
of
title
there.
B
Scientists
can
log
in
they
can
use
notebooks,
they
can
do
their
experimentation,
they
can
deploy
it
into
production
for
inferencing
and
they
can
bring
it
into
application
pipelines.
So
so
what
they
did,
this
in
steps
and
every
good
solution,
I
personally
believe,
should
be
done
in
step,
so
for
the
first
step
was
to
create
a
self-service
cloud
like
experience.
So
this
is
think
about
as
the
situation
before
they
applied
apply.
B
The
solution
write
a
bunch
of
data
scientists
there
that
individual
machine,
some
of
them
may
even
have
GPUs,
and
they
may
have
bulky
machines
that
they're
not
sharing
this
they're,
not
they're
siloed,
they
don't
did
they're,
not
I
mean
the
data
that
they
access
is
bespoke.
They
can't
share
computer
resources
so
to
solve
that.
What
do
they
do?
B
They
put
that
all
those
nodes
in
a
cluster
in
a
queue
Banaras,
cluster
and
openshift
cluster,
and
they
put
jupiter
hub
jupiter
hub-
was
basically
their
self-service
portal
for
the
data
scientists
both
intra
for
interactive,
as
well
as
for
just
job
bad
jobs,
kind
of
projects
right
so
also
the
data
scientists
were
able
to
create
the
notebooks
do
their
experimentation.
All
that
I
try
to
work
that
I
explained
earlier,
and
so
they
were
able
to
achieve
this
as
the
first
step
in
the
journey.
B
You
know
so
one
of
the
you
saw
it
on
the
PA
side,
but
one
of
the
things
that
they
wanted
to
achieve
really
is
that
they
wanted
some
jobs
or
some
kinds
of
projects
required.
The
access
to
multiple
GPUs
inside
a
container,
whereas
some
others
they
wanted
to
share
GPUs,
so
they
had
both
ends
of
the
spectrum,
so
the
action.
So
this
is
a
way
of
me
showing
that
you
know
a
container
is
being
shared.
B
A
multiple
GPUs
are
attached
to
a
container
so
that
the
workload
running
in
them
can
take
advantage
of
these
multiple
GPUs,
whereas
the
next
one
is
really
about
how
you
share
a
given.
Gpu
GPU
has
a
lot
of
CUDA
cores
and
memory.
So
how
do
you
share
it
among
multiple
containers
right,
so
they
they
wrote
something
actually
which
was
internal
to
make
this
happen.
B
B
How
we
do
that
correctly,
but
you
basically
establish
bunch
of
queues
and
so
jobs
are
coming
in
and
if
it,
the
cuban
Eddie
scaler,
knows
how
to
basically
take
up
an
entire
GPU
and
assign
it,
and
this
works,
because
all
those
those
machine
learning
algorithms
that
power-hungry
right
remember,
we
talked
about
that
so
they'll
use
it
all
hundred
percent
anyways,
and
so
that
should
be
ok.
So
so
the
next
step
they
did
was
multi-tenancy.
Since
there
were
a
number
of
projects,
a
number
of
data
scientists
they
needed
to
do.
B
Multi-Tenancy
and
again,
we'll
start
with
this:
a
bunch
of
cluster
with
a
bunch
of
nodes
and
the
jib
that
have
running
on
it.
And
then
the
next
step
you
do
is
that
you
create
this.
As
an
illustration,
you
create
a
project
called
echo
and
using
kubernetes
namespaces.
Then
you
create
a
project
called
tango,
and
then
you,
data
scientists
on
this
project
are
able
to
access
data
sets
in
separate
storage
classes
and
PVCs,
so
that
you
know
data
scientists
from
this
project.
B
Echo
are
able
to
only
access
data
there
and
are
able
to
do
their
modeling
and
experimentation
and
data
scientists
from
project
tango
are
able
to
access
only
data
and
resources
on
that
side.
So
you
know,
and
I
won't
go
through
each
one
of
them,
because
this
is
cube
con
and
you
probably
are
already
familiar
with
some
of
the
ways
to
make
this
happen.
B
Using
our
back
is
a
network
policy
is
using
storage
classes
per
project
or
a
namespace,
and
then
resource
isolation
with
everything
from
C
groups,
all
the
way
up
to
resource
manager
and
also
having
resource
code.
As
so
you
take
that
one
step
further.
You
want
to
be
good
neighbors
with
each
other.
Again,
you
start
you
you
you
so
the
first
thing
you
do
is
that
you're,
you
GPUs
are
precious,
so
you
don't
want
some
arbitrary
jobs,
which
is
only
GPUs
to
land.
On
that
so
use
Tencent
relations.
B
B
You
know
jobs
or
parts,
or
what
have
you
with
you
know
with
the
toleration
for
these
and
and
that
would
be
kind
of
a
hard
way
of
doing
it,
whereas
if
you
want
to
use
a
more
software
way
to
do
that,
you
could
use
affinity
and
anti
affinity
and
then
now,
obviously,
if
you
do
not
want
your
data,
scientists
to
you,
know,
write
it
in
the
ports
back.
Then
you
can
create
an
admission
controller.
B
I
you
know
called
e,
for
example,
the
GPI
admission
controller,
which
adds
all
these
you
know
to
the
part
before
it
lands
to
the
scheduler,
and
then
you
can
do
multi.
You
can
do
other
things
such
as
you
can
create
multi
projects.
We
saw
it
on
the
previous
page
and
each
project
can
have
quotas
for
GPU
usage,
so
there
are
ways
to
be
good
neighbors
and
they
did
as
a
journey.
They
did
that
with
a
POC
and
then
as
they
get
into
production,
they
implemented
some
of
these
features.
B
So
so
far,
what
we
have
done
really
is
that
we
have
created.
If
you
go
back,
you
know
they
have
created
the
self-service
cloud-like
experience
for
the
data
scientist
right.
So
the
next
step
really
was.
You
know
what
is
known
as
ml,
which
I'll
describe,
but
the
problem
that
they
were
trying
to
solve.
Really.
Is
that
you
know
we
all
know
this,
but
data
scientists
are
not
that
talent
is
not
easy
to
get
right,
so
what
they
were
really
wanted
to
do
is
they
wanted
to
be
able
to?
B
Obviously
there
are
data
scientists
who
know
what
they're
doing,
but
then
they
wanted
to
kind
of
spread
this
talent
and
for
others
were
not,
maybe
so
who
may
not
be
subject
matter
experts
to
become
these,
so
I
think
the
the
premise
really
is
what
if
I
told
you
that
you
can
be
a
ml
expert
right,
so
that's
kind
of
the
premise,
and
so
what
did
they
do?
They
introduced
Auto
ml.
So
what
is
Auto
ml
right?
So
what
do
you
want
is
a
model.
B
You
want
some
kind
of
a
modular
set
of
models
right,
and
we
all
know
that
you
start
with
data,
and
then
you
have
something
that
you
want
to
optimize
right
and
then
you
have
some
set
of
constraints
that
you
you
have
for
the
problem
and
then
normally
what
you
would
have
done
really
is
that
you
would
have
done
this.
Some
data
scientists
would
have
done
would
have
picked
the
right
machine
learning
model
and
then
all
the
things
that
we
talked
about
earlier.
B
But
what,
if
we
can
automate
it
right
so
that
we
can
and
that's
basically
what
they
did
right.
So
this
is
the
auto
ml
is
a
platform
that
basically
automates
the
and
I
automates
the
forgiving
set
of
data
and
given
set
of
optimization
given
set
of
constraints.
It
tries
different
kinds
of
machine
learning,
models
and
I
treat
said
and
then
produces
something
that
is
at
the
desired
level
of
accuracy,
within
the
constraints
that
you
are
specified
so
as
an
illustration
it
I
actually
just
did.
B
B
Running
on
top
of
OpenShift
here
you
can
see
the
scikit-learn
part
running
there
and
then
jupiter
hub
itself,
and
you
know-
and
then
I
mean
obviously
we
all
know
this
I
mean
this
is
Cuban
Eddie,
so
you
can
scale
up
and
scale
down.
You
know
the
number
of
parts
you
can
use,
HP
or
horizontal
port
or
the
scale
at
Exeter,
etc.
So
we
can.
We
can
have
fun
with
this
for
sure.
So
what
are
some
of
the
design
elements
that
they
choose?
B
They
used
PostgreSQL
for
some
metadata
and
for
as
a
catalog,
and
then
they
use
rabbitmq
for
as
the
message
queue
they
use
floss
to
deploy
those
models
into
as
restful
services,
and
all
this
was
running
on
top
of
open
shelves.
Sorry,
so
what
are
some
of
the
results
that
they
saw
right?
So
this
is
ultimately
what
we
somehow
care
for
here
about
right.
So
this
was
a
30
percent
average
implement
in
machine
learning
model
performance,
a
70%
increase
in
machine
learning
models.
Right
again,
this
did
not
require
that
they
scale
the
number
of
data
scientists.
B
This
is
because
you
know
you
know
non
data.
Scientists
could
take
advantage
of
the
auto
ml
platform
and
they
could
create
these
services
based
on
machine
learning
and
then
exclusive
growth
in
a
number
of
users.
You
know
using
this
platform,
so
that's
all
great.
So
what
are
some
of
the
challenges?
And
next
steps
that
you
know
they
are
seeing
the
first
step
really
is
you
know,
automate
development,
so
one
of
the
things
is
that
right
now,
this
this
thing
is
great,
but
they
wanted
automation.
B
They
wanted
to
automate
the
end-to-end,
workflow
and
I'll
describe
what
that
means
in
a
later
slide,
because
we
are
helping
them
with
all
that.
You
know
they
wanted
a
better
way
to
say
when
you
know,
obviously,
all
the
models
were
saved
but
better
way
to
catalog
the
experiments
they
wanted
to
extend
this
auto
ml,
the
to
unstructured
data
like
images
and
audio.
They
want
to
again
have
a
better
and
supported
way
to
do.
B
Gpu
sharing
and
obviously
multi
clusters,
as
this
gets
more
and
more
popular,
they
want
be
able
to
deploy
more
and
more
clusters,
and
so
my
cluster
is
really
important
for
them.
So
what's
next-
and
this
really
is
the
collaborative
approach
that
we
have
from
a
red
head
perspective,
how
we're
working
with
them
to
realize
and
these
some
of
these
challenges.
The
first
thing
is
obviously
this
Red
Hat,
so
we're
always
community
first,
so
we're
working
with
the
communities
and
I
won't
go
to
each
one
of
these.
B
Obviously,
but
we
are
working
with
all
the
relevant
machine
learning
and
AI
communities.
A
basic
you
flow
is
obviously
the
big
one
there
and
we
are
also
working
with
some
of
the
other
ancillary
in
that
context,
communities
like
Kafka
and
and
Cassandra
all
these
other
things
that
you
need
really
to
create
an
entire
end-to-end
system
that
we
saw
earlier
and
we
are
working
with,
for
example,
Nvidia
and
GC.
So
what
we
are
doing
is
we
have
this
operator
hub,
dot,
IO
and
I'll
describe
what
that
is
the
next
page.
B
But
basically,
this
is
a
community
based
website,
for
you
know
where
you
can
find
operators
for
each
one
of
these
and
more
right,
and
some
of
them
might
be
in
queue
flow.
For
example,
the
PI
torture,
the
tensorflow
operator
might
be
in
q4,
that's
okay,
but
that's
kind
of
where
we
are
today
and
then
the
other
thing
that
we're
doing
is
that
we
have
this
reference
architecture
called
open
data
hub
again.
This
is
upstream
in
the
community
and
that
offers
how
to
build
a
a
machine.
Learning
is
a
service
kind
of
similar.
B
What
you
saw
earlier
on
openshift,
so
so
this
is
basically
how
I
operate.
Have
dot.
Io
looks
like
you
can
see.
There
are
operators
for
all
kinds
of
you
know,
tools
and
software
here,
I
won't
go
into
each
one
of
them.
Open
data
hub
itself,
I'll
describe
next.
So
what
it
is
is
the
open
data
hub
really
is
that
reference
architecture?
What
it
can
do
today
is
it
can
deploy
all
these
components
at
like
spark
and
Kafka
for
streaming.
B
Your
data
into
openshift
and
into
memory
Jupiter
hub
is
a
self-service
portal
for
data
scientists
for
notebooks.
Both
interactive
and
otherwise
there
is
cell
for
storage
that
is
Seldon
4n
trancing.
This
comes
from
to
flow,
and
then
there
is
griffons
permutation
graph
Anna
to
do
to
do
the
tracking,
the
monitoring
that
I
mentioned
earlier
about
drift,
etc,
and
all
this
is
deployed
by
open
data
hub
with
this
operator
called
the
open
data
hub
operator,
which
then
manages
the
end-to-end
life
cycle
of
all
this.
B
So
that's
one
of
the
advantages
that
the
operator
have
has
our
operators
in
general
operator
lifecycle
manager
has
is
that
they
can
manage
the
lifecycle.
If
you
want
to
update
the
apache
spark
due
to
whatever
CBE
what-have-you
or
there's
a
fix
or
a
new
feature
work,
you
know
you
can
update
it
using
the
operator
concept
right.
B
So
what
else
have
we
doing
open
ship
for
itself
we're
continuing
to?
Obviously
in
Western
we
have
a
robust
roadmap.
You
know
GPU
enabled
that's
already
there
nothing
new
there,
but
then
we
also
have
open
shift
service
mesh
with
Sto
and
Egger
and
Prometheus
and
key
Ali.
Then
we
have,
you
know
key
native
server
list
with
key
native
and
Tecton,
and
then
all
this
can
be
basically
deployed
with
the
openshift
installer
on
various
infrastructure.
B
So
this
is
where
they
can
scale
right,
I
bring
it
up
only
because
I
want
to
show
that
how
they
can
scale
as
they
want
to
do
more
clusters.
One
of
the
fundamental
aspects
is
this:
so
how
does
this
look
like
from
experimentation
to
production
with
CI
CD?
So
this
is
the
other
thing
ever
we
said
they
wanted
to
do
more
automation.
So
how
does
how
do
you?
How
do
you
automate
this
end-to-end
workflow
for
the
data
scientists?
This
is
just
an
illustration.
This
is
a
data
scientist
with
the
notebook.
B
You
know
they
check
it
into
associate
processors
get.
Then
you
know
you
the
source
to
image,
basically
converts
this
notebook
into
a
container
image.
It
then
gets
deployed
on
the
openshift,
you
know
and
then
they
you
know
I
tration
can
happen.
I
mean
the
notebook
runs
with
whatever
compute
resources
and
data
it
needs,
and
then
there
is
enough.
B
You
know
accuracy
of
the
model
and
you
know
you're
satisfied.
It
can
then
go
into
production
right,
so
you
can
promote
and
serve
the
model
in
production.
This
is
something
that
we
already
do
for
just
vanilla
applications.
So
why
not
do
this?
For
you
know,
the
data
science
pipeline
is
what
this
is
saying
and
then
you
are
continuously
monitoring
this
for
performance
and
driven.
You
saw
of
the
open
data
hub,
actually
deploys
Prometheus
and
Griffin
er
to
make
that
happen.
B
So
that's
the
feedback
loop
and
that's
how
we
create
the
III
trade
you
loop,
so
with
that
I
thought
I
would
not
actually
reach
the
demo.
I
thought
I
would
mumble
along
at
some
point,
and
you
know
I
slow
myself
down,
but
luckily
I
have
so
I
will
be
able
to
show
this
demo
to
you
guys.
The
demo
is
really
it's
kind
of
rather
simple,
but
it
shows
how
open
data
hub
it
can
be
used
to
deploy
all
those
components
that
I
just
described
on
the
open
ship
cluster.
B
So
it's
it's
rather
simplistic,
but
I
think
it.
It
makes
the
point.
So,
let's
start
with
this,
this
is
the
open
shift
console.
Some
of
you
might
be
familiar
with
it,
and
what
you
do
is
that
you
start
but
I'll
roll.
This
and
I'll
pause
it
at
some
points
and
then
I'll
explain
what's
going
on
so
the
first
thing
really
is
that
you
know
as
a
admin.
B
You
know
this
is
you
know
you
are
the
admin
and
you
are
managing
this
cluster.
You
can
create
a
project
in
just
a
few
seconds
if
it
is
ruling
you'll
see
that
as
an
admin
you
have
logged
in,
you
will
create
a
project.
You
know,
I
have
called
the
open
data
hub
project,
you'll,
give
it
a
name
and
the
next
step
after
that
is
that
you
will
want
to
now
install
open
data
hub.
B
You
know
you
can
browse
down
to
this
call
think
of
the
embedded
operator
hub,
and
that,
basically,
is
has
all
the
stuff
that
you
saw
on
operated
hub,
dot,
IO,
and
then
you
can
look
for
you
know
open
data
hub
operator,
and
then
you
can
install
it.
You
know
you
can
see
that
that's
happening
here.
It
just
says
that
it's
a
community
operator.
This
is
not
something
that
is
certified.
B
It's
not
a
quote,
unquote
product.
It
has
some
more
information
here
on
what
can
be
done
with
it.
It
shows
all
the
things
that
it
can
deploy
the
whole
list,
graph,
Anna
and
2,
Sigma,
AI,
librarian,
Argo
and
Selden,
etc,
etc.
But
in
this
particular
in
this
particular
demo,
we're
just
going
to
install
Jupiter
event
spark
for
the
purposes
here
is
where
that
is
happening,
which
I
mentioned
earlier.
You
can
create
an
operator
subscription
and
you
can
do
it
to
a
specific
name
space.
In
this
case.
B
The
open
data
hub,
which
you
just
created
earlier
following
that
what's
going
to
happen,
is
that
you
know
the
the
openshift.
The
controller
is
going
to
create
this
open
shift
or
the
open
data
hub
operator
I'll
speed
this
along
a
little
bit,
because
you
know
you
can
see
that
all
those
being
created
now
you
can
see
that
the
opener
to
have
operator
has
been
installed.
B
In
this
particular
case,
as
I
said,
we
are
going
to
not
deploy
most
of
it,
except
for
Jupiter
hub
and
spark,
and
so
that's
basically
what
is
happening.
We
still
have
five
minutes
I'm
going
to
let
it
roll,
but
we
can
have
also
some
questions
later.
So,
let's
see
what's
going
on
here
yeah.
So
this
is
just
illustrating
the
fact
that
these
are
the
radius
I'll,
not
speak
for
a
minute,
so
that
you
can
absorb
it.
B
Monitoring
is
not
being
deployed,
like
I,
said,
I,
think
this
part
and
and
then
you
can
modify
the
amount
of
CPU
and
memory
that
you
want
for
that
operator
and
then
I'll
speed
it
along
here
a
little
bit.
You
know.
Obviously
it's
editing
Hamill
files
is
not
that
interesting,
but
you
get
the
idea,
you
know,
and
then
you
know
you
create
it
now.
You
now
the
example
open
data
hub
has
been
created.
B
Now
you
can
go
to
a
ports
and
you
can
see
that
Jupiter
hub
has
been
deployed
and
then
you
can
see
that
there
is
the
SPARC
operator
down
below
so
now
it
goes
into
now.
How
are
you
going
to
use
it
so
it'll
go
into
routes
in
just
a
moment,
and
then
you
know
a
route
is
created
to
the
Jupiter
hub,
and
so
you
are
you
going
to
use
that
you're
going
to
click
on
that
and
then
you
are
able
to
login
into
the
Jupiter
hub.
Just
second,
you
can
sign
it.
B
So
that's
the
other
nice
thing
you
don't
have
to
sign
it
again
and
again.
You
can
sign
it
with
OpenShift
credentials
that
you
are
previously
logged
in
already
I.
Believe,
that's
what
you
can
do.
So,
let's
keep
going,
and
so
this
is
what
you
get
is.
This
is
the
familiar
spawner
that
you
get
with
Jupiter
hub
here.
As
you
can
see,
you
can
pick
from
some
pre-canned
notebooks.
This
one
happens
to
be
a
spark
minimal,
spark
notebook
operated
there.
You
can
see
that
you
can
pick
different
pre-canned
container
sizes.
B
You
can
say
this
much
which
have
so
much
memory,
CPU
and
memory.
You
can
pick
the
number
of
GPUs
you
want
for
this
and
then
you
can
also
add
an
ornamental
variable.
Sister's,
a
tableau
as
access
key
ID
is
etc
for
the
data
part
of
it.
In
this
particular
case,
we
don't
use
it
because
we
are
not
using
Saif
s3
in
this
particular
demo,
but
that's
something
that
you
can
do
so.
The
next
thing
you
do.
B
You
respond
the
the
the
SPARC
operator,
and
this
basically
does
a
simple
and
then
you
in
this
particular
case
you
are
going
to
get
a
simple
notebook
which
is
empty,
and
then
this
just
shows
you
you
can
upload
your
own
notebooks
and
create
a
empty
notebook,
and
this
is
just
a
simple
hello
world
anyways.
So
that's
kind
of
really
what
I
wanted
to
show
in
terms
of
a
simple
demo
on
what
happens
and
one
we
could
go
back
to
my
presentation
to
conclude.
B
To
conclude
right,
so
in
summary,
what
we
saw
today
was
the
case
study
for
a
Modi.
How
did
they
build
a
machine?
Learning
is
a
service
platform
using
open
shift
and
other
open
source
technologies.
You
know
what
we
saw
why
they
needed
to
build
it
and
how
they
built
a
cloud
like
experience.
We
saw
how
they
built
auto
ml.
We
saw
how
cuban
IDs
and
open
other
open
shift
tools
can
help
with
this
and
they're.
B
In
today
I
mean
we
don't
have
to
wait
for
anything
and
then
finally,
we
saw
what
the
operator
have
in
the
open
data
hub
do
so
with
that.
I
have
conclude:
these
are
my
contact
information
and
we
can
open
up
for
some
Q&A
I'm
be
around,
and
so
obviously,
if
we're
kicked
out,
we
can
answer
questions
outside.