►
From YouTube: Keynote: Machine Learning on Kubernetes Made Easy With Kubeflow - Masoud Mirmomeni & Jimmy Guerrero
Description
Don’t miss out! Join us at our next event: KubeCon + CloudNativeCon Europe 2022 in Valencia, Spain from May 17-20. 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.
Keynote: Machine Learning on Kubernetes Made Easy With Kubeflow - Masoud Mirmomeni, Lead Data Scientist, Shell & Jimmy Guerrero, Vice President of Marketing, Arrikto
B
A
Hi,
my
name
is
jimmy
and
I
run
developer
relations
at
ripto.
For
those
of
you
who
may
not
be
familiar
with
the
ricto,
a
richto
was
a
key
contributor
to
the
kubeflow
1.3,
as
well
as
the
recently
cut
1.4
release,
despite
or
besides,
participating
in
couple
in
a
couple
of
the
keyflow
working
groups,
they're
also
the
primary
maintainers
of
the
mini
kf,
as
well
as
the
kl
projects
plus
ekf,
which
is
the
enterprise
kubeflow
ml
ops
platform.
A
Now
I
know
that
we're
at
cubecon,
so
this
is
going
to
be
a
conference.
It's
all
about
cloud
native
architectures,
not
necessarily
data
science
or
machine
learning,
which
is
to
say
that
likely,
the
majority
of
the
audience
or
the
viewers
that
are
joining
us
virtually
are
going
to
be
cloud
native
developers
or
architects,
not
necessarily
ml
ops,
practitioners
or
data
scientists.
So
it's
not
going
to
hurt
to
spend
just
a
minute
to
talk
about
why
the
combination
of
kubernetes
and
machine
learning
are
actually
a
match
made
in
heaven.
Here's
why?
A
A
So
these
are
going
to
be
things
that
are
going
to
be
doing:
data
management,
security,
maybe
front
end,
visualizations,
etc,
and
here
we're
going
to
probably
want
to
go
with
a
microservices-based
container
architecture.
Here
again,
kubernetes
is
going
to
be
a
slam
dunk
for
us.
Finally,
machine
learning
loves
gpus,
but
gpus
are
expensive
right,
so
it's
not
always
about
how
quickly
can
we
spin
up
an
environment
and
get
access
to
all
the
resources
that
we
need?
A
In
this
case,
it
could
be
just
as
important
how
quickly
we
can
spin
down
that
environment
back
down
to
zero.
So
here
again,
containers
are
going
to
be
a
perfect
fit.
Unfortunately,
there's
an
open
secret
in
in
in
the
industry
that
not
a
lot
of
or
a
a
lot
of
machine
learning
models
are
not
being
suspect,
successful
in
making
it
to
production,
and
the
question
is:
why
is
this
well?
A
Therefore,
finding
the
right
methodology,
the
right
software
and
perhaps
a
little
bit
of
empathy
right,
that's
going
to
be
needed
in
order
to
collaborate
about
across
these
teams
and
be
successful,
can
prove
to
be
a
little
bit
elusive.
So
what
are
we
to
do?
Enter
kubeflow
kubeflow
is
the
open
source
project
smackdab
in
the
middle
of
this
big
convergence
in
it?
And
here
I'm
talking
specifically
about
the
combined
ubiquity
of
cloud
native
architectures
and
the
needs
of
machine
learning
workflows.
A
As
we
know,
kubeflow
was
originally
launched
by
google
back
in
2017
and
has
since
become
the
most
robust,
open
source
cloud
native
by
design,
not
as
an
afterthought
ml
platform
for
data
scientists
as
well
as
operations.
Folks,
it's
a
complete
toolkit
of
components
that
allow
both
data
scientists
and
operators
to
manage
train
model
and
tune
and
even
monitor
their
workflows.
A
Now
that
I've
said
a
little
bit
of
context,
I'm
going
to
hand
it
back
over
to
massoud
who's
going
to
walk
us
through
part
of
shell's
data
science
and
machine
learning
journey.
So
we
can
understand
how
q
flow
and
its
ecosystem
of
integrations
helped
solve
many
of
the
challenges
that
they
were
facing
massoud
over
to
you.
B
Thank
you
most
of
you
might
know
shell
as
the
old
giant.
However,
in
these
recent
years,
shell
has
expanded
its
focus
and
and
to
other
sorts
of
energies,
are
green
and
renewable,
and
to
that
effort
actually
roughly
spent
two
billion
dollars
annually
through
2020
for
these
kind
of
new
resources
and
expected
to
expand
these
expenses
to
even
more
for
years
to
come.
B
So,
like
you
know,
it's,
it's
obvious
stepping
into
these
very,
very
large-scale
environment
that
you
need
to
get
your
resources
from
different
source
of
energy
and
distribute
and
transmit
to
users
that
are
increasing
day
by
day
and
they
have
drastically
different.
You
know:
consumption
patterns
need
a
smart,
very
fast,
agile
control
system
and
without
having
artificial
intelligence
at
the
scale.
This
is
not
gonna
be
achievable,
but
having
ai
at
shells,
the
scale
can
create
some
challenges
and
our
team
at
shell
faced
some
of
these
challenges.
B
The
first
challenge
that
we
had
was
creating
a
development
environments,
proper
development
environments
for
these
kind
of
challenge
problems.
So,
as
a
data
scientist,
I
used
to
work
on
the
local
environments,
like
you
know,
build
some
machine,
learning,
simple
machine
learning
models
and
using
the
local
data.
Now
we
are
going
to
use
a
large
scale
data
set
from
the
greece
from
different
countries
all
around
the
us
europe,
and
we
want
to
build
the
model.
B
So
it's
going
to
be
extremely
hard.
If
you
want
to
create
an
environment
like
that
for
to
be,
you
know
capable
of
doing
some,
some
modeling
like
that.
Well,
the
second
challenge
we
want
to
work
in
these
environments.
These
environments
require
the
specialized
skills.
Okay,
if
you
want
to
work
on
your
local
machine
as
a
simple
model,
it's
going
to
be
really
really
easy,
but
you
want
when
you
want
to
go
and
grab
this
data
set.
Let's
say
you
want
to
forecast
price,
you
want
to
create
load
consumption.
B
You
want
to
figure
out
whether
you
want
to
figure
out
the
generation.
Is
the
consistency
in
the
grids
you
need
to
have
a
graph
of
your
network
and
you
need
to
combine
and
get
the
data
it's
going
to
be
extremely
hard,
and
now
you
want
to
run
them
on
kubernetes.
This
is
great,
but
before
that
you
need
to
know
about
containers,
you
need
to
about
the
how
to
scale
you
need
to
know
about
gpus
before
even
getting
to
the
modeling.
This
can
take
very
very
long
time.
B
B
I
wish
we
cannot
give
any
a
couple
of
gpus
to
every
data
scientist
on
top
of
that
machine
learning
is
a
very
sparky
process
when
you
are
in
the
development
environment,
when
your
code
is
ready
to
be,
and
it's
in
production,
you
just
need
a
couple
of
cpus
to
have
it
running.
But
when
you
are
in
the
modeling
phase
and
as
I
mentioned,
that
the
problem
is
very
huge,
so
you
need
to
have
a
very
huge
search
space.
You
need
to
tune
so
many
parameters
and
you
need
to
have
a
huge
computational
power.
B
If
you
are
in
the
production
environment
and
you
have
a
huge
computational
power,
you
lose
it
money
because
you
don't
need
that.
On
the
other
hand,
when
you
are
in
the
modeling
phase,
you
want
to
have
a
huge
computational
power,
because
if
you
don't
have
it
you're
going
to
lose
money
by
wasting
the
expensive
time
of
your
data
scientist
and
now
we
want
to
see
how
kubeflow
actually
help
us
to
address
all
of
these
challenges.
B
The
first
thing
is
going
to
be:
the
tupelo
actually
creates
a
self-serving
model
for
us,
so
you
and
data
scientists
can
go
and
grab
computational
power
and
storage
and
they
have
pre-configured
ml
toolkits
in
that
that
is
that
exists
in
the
secure
cloud
environments.
How
cool
is
that
now
we
can
actually
bring
all
of
those
things
and
do
the
machine
learning
projects
easily.
You
know
from
minute
zero.
B
The
second
one
with
having
q
flow,
automated
pipeline
engine
or
say
mkhk.
We
are
going
to
fill
in
the
gap
between
the
data,
science
and
software
engineering
and
mlrs.
Now
our
data
scientists
are
capable
of
using
their
simple
code
and
and
bring
it
and
pass
it
to
the
mla
so,
and
this
makes
the
process
much
faster
for
us
to
put
things
into
production.
And
finally,
since
we
are
using
kubernetes,
we
can
smartly
manage
our
computational
and
storage
resources.
B
B
We
just
need
to
give
a
name,
let's
say
cubecon
and
you
can
see.
We
have
jupyter
notebook
environment,
we
have
visual
studio,
r
studio
and-
and
if
you
remember
I
mentioned,
we
have
a
different
ml
toolkit
pre-configured
ml
toolkit,
you
have
something
for
deep
learning,
different
version
of
tensorflow
pytorch.
We
have
something
for
spark.
We
have
gpu
version
of
that
and
if
there
is
something
that
doesn't
exist
here,
it's
easy
to
actually
bring
it
up
here
for
other
applications.
B
After
that,
with
some
simple
configuration
we
are
ready
to
go.
We
just
need
to
say
how
many
cpus
I
need
how
much
memory
I
need
to
have
for
my
server
and
if
I
need
gpu
or
not,
for
example
here
I
just
I
don't
need
gpu
for
the
simplicity
and
after
that,
I'm
just
going
to
say
how
much
storage
hard
I
need
for
my
notebook
server
and
it's
I'm
going
to
skip
like
you
know,
for
the
simplicity
for
the
some
other
configuration
and
we're
ready
to
go.
B
Just
click
on
this
beautiful
launch
button
and
you're
going
to
see
my
notebook
survey
is
going
to
start
in
couple
of
seconds
which
could
take
me
like
a
couple
of
months.
You
know
without
having
these
things
now
we
are
ready
to
connect
and,
as
you
can
see,
I
just
need
to
have
a
web
browser
and
a
secure
internet
connection.
Now
I
am
in
my
server,
I
have
visual
studio.
B
I
have
jupyter
lab
and
now,
if
you
go
to
the
jupiter
lab
you're,
going
to
see
it's
very
similar
to
our
lovely
jupiter,
notebook
and
but
there's
something
more
to
that,
we
are,
we
have
secularly
connected
to
aws
and
all
of
my
data
is
located
there.
So
I
can
bring
and
drag
everything
to
my
jupiter
notebook
and
I
can
start
doing
some
data
science
and
cool
stuff
from
minute
zero.
B
I'm
going
to
suffer
a
little
bit
in
this
graph
and
I'm
going
to
share
the
beauty
I
see
in
this
graph.
For
you,
it's
just
some
might
be
a
very
simple
flow
graph,
but
this
graph
was
very,
very
lucky
lovely
to
me.
It
gave
me
those
one
of
those
aha
moments
when
I
saw
it
for
the
first
time.
I
was
super
excited
when
I
joined
shell
as
a
data
scientist.
My
first
assignment
was
to
build
a
predictive
model.
B
I
needed
to
grab
data
from
different
sources,
I
needed
to
subsample
them
and
I
needed
to
use
different
model
configuration,
but
I
couldn't
have
a
huge
search,
especially
because
I
was
working
locally,
long
story
short.
It
took
me
a
month
or
two
months
actually
to
come
up
with
the
like
in
the
modeling,
the
proof
of
concept
and
jupiter
format.
I
pass
it
to
my
co-worker
and
I
said:
can
you
productionize
that
it
took
him
a
month
to
come
up
with
a
model
and
the
performance
was
not
that
great?
B
I
was
lucky
at
that
time
to
be
accounted
to
eric
and
queue
flow.
With
the
help
of
my
co-worker,
we
built
the
machine,
learning
discipline
and
repeated
the
same
experiment
in
just
35
days
from
the
data
processing
to
deployment
and
exponentially
reduced
the
timing.
Our
second
effort
to
couple
of
days
now
in
our
team,
we
have
some
team
members
with
basic
programming
skills
that
can
apply
cutting-edge,
machine
learning
and
deep
learning
in
just
a
couple
of
hours,
and
the
story
doesn't
end
here.
It's
getting
get
even
simpler
and
better
for
data
scientists.
B
We
data
scientists,
love
jupiter,
no,
what's
going
on
jupiter
notebook
now
we
just
grab
this
jupyter
notebook,
add
some
add
some
text
to
those
cells
like
import
pipeline
skip
and
some
others,
and
we
can
push
a
button
and
create
a
pipeline
from
it.
Kale
is
going
to
take
over
that
code
for
us
and
create
a
valid
pipeline,
and
it's
going
to
take
care
of
all
the
data
dependencies
and
it's
going
to
manage
the
life
cycle
of
this
cube
flow
pipeline.
B
And,
of
course,
finally,
snapshot
policies
allow
us
to
release
idle
resources
without
losing
any
work,
and
these,
ladies
and
gentlemen,
was
the
game-changing
experience.
I
wanted
to
share
with
you
as
a
data
scientist
actually
and
how
kubeflow
actually
helped
me
to
focus
on
my
work,
and
you
know
avoid
all
of
those
distractions
that
I
was
always
hesitating
to
touch
that.
So
I
could
focus
on
my
work
and
challenges
that
we
have
in
huge
projects
that
we
have
at
shell
and
be
productive
and
deliver
the
projects
in
a
timely
manner.