►
Description
Want to view more sessions and keep the conversations going? Join us for KubeCon + CloudNativeCon North America in Seattle, December 11 - 13, 2018 (http://bit.ly/KCCNCNA18) or in Shanghai, November 14-15 (http://bit.ly/kccncchina18).
Keynote: Kubeflow ML on Kubernetes - David Aronchick, Product Manager, Cloud AI and Co-Founder of Kubeflow, Google & Vishnu Kannan, Sr. Software Engineer, Google
About David
David Aronchick was the Senior Product Manager for the Google Container Engine and led product management on behalf of Google for Kubernetes. David has been helping to ship software for nearly 20 years, founding and being part of the management team for three different startups, as well as squeezing in time at Microsoft, Amazon, Chef, and now Google. David is co-founder of the Kubeflow project, an effort to help developers and enterprises deploy and use ML cloud-natively everywhere.
About Vishnu
Vishnu Kannan is a Senior Software Engineer at Google. Vishnu received his Masters in ECE from Georgia Tech. He has been a systems engineer ever since he graduated. He hacked on the Linux Kernel for a couple of years at Cisco. He then worked on Borg at Google. He is currently an active maintainer in the Kubernetes community focussing on managing compute resources at scale."
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
So
I
think,
let's
skip
past
ml
to
get
to
the
really
important
bits.
How
are
you
all
doing
on
cube
con
bingo?
Do
I
need
to
say
something
make
a
joke
about
Maersk
or
something
to
get
someone
a
bingo
happy
to
do
it.
I'm
Dave,
as
Kelsey
just
said,
he's
vish
we've
been
at
every
cube.
Con
ever
I
feel
like
we're
cute
con
hipsters.
We
were
here
before
it
was
cool,
but
you
know
thinking
back
to
those
days
of
the
original
kubernetes
and
cube
con
and
things
like
that.
A
It's
crazy
to
think
about
how
many
ways
there
were
to
run
containers
cron,
tabs
you're,
an
Orchestrator
bash.
Looking
at
you
up
and
shift
on
bash
but
honest
to
god,
it
was
just
everything
was
bespoke.
You
wrote
it
yourself,
you
ran
it
yourself.
You
had
to
deal
with
everything
yourself,
but
a
couple
days
ago,
Alexis
Richardson
talked
about
the
transformation
that
kubernetes
brought,
because
it
gave
everyone
a
common
platform
that
they
could
trust.
A
A
This
is
what
people
see
when
they
see
ml.
This
is
Google's
public
information
about
our
data
centers.
This
is
the
best
data
center
engineers
in
the
world.
These
are
people
who
spent
their
entire
lives
getting
great
at
power
usage
effectiveness
they
these
are
people
are
using
renewable
they're
using
all
the
latest
technologies
they
attach
ml
to
it
and
they
save
40%
hundreds
of
millions
of
dollars
a
year,
and
you
know
on
first
blush.
They
didn't
even
know
why?
Because
no
one
can
understand
what
ml
model
do.
Does
that's
the
power
of
ml?
A
It
is
great
and
it
is
very,
very
hard
and
it's
not
just
hard,
because
I'm
bad
at
math,
it's
hard,
because
people
see
this
ml
goodness,
but
most
folks
are
way
way
over
there
and
in
between
you
have
lots
of
pain,
and
where
does
the
pain
come
from?
Well,
it's
funny
because
it
comes
from
the
same
place
it
that
we
were
with
kubernetes.
You
know
three
years
ago.
First,
you
have
basically
two
options,
one
you
could
do
it
yourself,
meaning
set
it
up
from
scratch.
You're
great
with
legacy
systems.
A
You
eventually
get
to
some
business
need
and
get
something
working
and
then
you
have
to
migrate,
and
then
you
have
to
change
based
on
business
need,
and
you
have
to
do
it
all
over
again
and
it's
an
incredible
pain
in
the
butt
or
you
could
go
with
hosted
and
a
host.
It
is
great
for
five
minutes
and
then
the
next
five
years
of
your
life,
you
spend
doing
everything
you
did
before
you
have
to
take
it
apart.
You
have
to
figure
out
how
to
integrate
with
your
legacy
systems
and
by
the
way
you're
locked
in.
A
A
So
we
need
cloud
native
ml
and
cloud
native
ml
would
do
this
exact
thing
we
think
of
cloud
native
ml
as
really
adopting
the
following
three
components:
composability
Portability
and
scalability.
By
compile
possible
I
mean
today
you
build
a
model,
and
this
is
where
everyone
focuses
their
time.
But
this
is
not
where
data
scientists
focus
their
time.
They
end
up
focusing
their
time
on
everything
built
around
this
a
whole
set
of
microservices
loosely
coupled
together.
That
together
comprise
what
is
really
an
m/l
solution.
A
All
the
way
from
that
initial
thought
bringing
your
data
in
splitting
it
training
it
production,
izing
it
and
then
rolling
it
out
to
production.
That's
what
an
ml
solution
looks
like
it
does
not
look
like
just
building
a
model
and
worse,
every
single
data
scientist
is
gonna,
have
a
slightly
different
view
on
what
the
right
tools
are
and
how
to
actually
roll
these
things
out,
and
that
means
you
need
a
highly
composable
solution,
one
where
every
data
scientist
can
choose
the
components
that
are
meaningful
for
them.
Second,
you
want
it
to
be.
A
Portable
I
already
talked
about
the
fact
that
that
every
data
scientist
and
every
solution
is
going
to
be
slightly
different,
and
this
is
what
a
stack
looks
like
and
I'm,
probably
leaving
off.
You
know,
100
other
boxes
here,
the
the
entire
stack
that
I
talked
about
earlier.
That
gets
into
just
that
top
part.
A
A
lot
of
people
think
about
scalability,
just
more
nodes
in
ml,
certainly
more
nodes
help,
but
you
also
have
to
think
about
all
these
other
elements,
accelerators,
CPUs
disk
and
then
how
about
some
non
Hardware
things
of
scale?
How
about
skill
sets?
Are
you?
Is
this
an
intern,
a
data
scientist,
a
Swee,
a
researcher?
How
about
teams
are
they
geo
distributed?
A
The
problem
is
that
data
scientists
are
really
really
good
at
doing
data
science,
but
they
don't
have
suite
or
IT
ops
in
their
name.
They
really
want
to
focus
on
what
they
do
well
and
that's
why
we
introduced
cube
flow.
We
want
to
make
it
easy
for
everyone
to
develop,
deploy
and
manage
portable
distributed
ml
and
kubernetes.
We
introduced
it
at
the
last
cube
con
and
we
have
been
so
overwhelmed
with
the
people
who
are
contributing
to
it.
A
The
idea
behind
cube
flow
is
to
try
and
address
all
three
of
those
components
simultaneously:
composability
scalability
and
portability.
We
leverage
the
fact
that
kubernetes
exists
everywhere.
It
operates
in
all
these
environments.
From
your
laptop
to
your
training
to
your
cloud,
you
create
your
configuration
out
of
cube
flow,
and
then
you
roll
it
out
into
each
location,
and
all
you
have
to
do
is
move
your
change,
your
parameters
in
order
to
adapt
and
we're
very
proud
today
to
announce
cube
flow
0.1
or,
if
you're
from
the
southern
hemisphere
1.0
in
the
Box.
A
Today
we
have
all
the
components
for
an
instant-on
ml
solution.
Did
I
mention
it's
0.1?
Please
don't
use
it
in
production,
yet
we
are
very,
very
excited
about
it,
but
it
is
right
at
the
beginning
and
still
you
are
able
to
get
very
productive
models
built
already.
You
have
Jupiter,
you
have
distributed
training,
you
have
model
serving
and
best
of
all.
A
You
have
packaging
to
let
you
customize
it
yourself,
without
making
changes
to
them
and
allowing
api's
without
doing
a
bunch
of
different
package
installs
from
different
operating
systems
and
distributions,
and
with
that
I'd
like
to
hand
it
off
to
vish,
to
show
a
demo
of
what
this
might
look
like.
So.
B
The
goal
here
is
to
show
the
operational
efficiency
that
you're
gonna
get
with
cube
flow
cameras
and
also
the
ease
of
use,
the
scalability
and
and
other
aspects
of
Hana.
So
it's
not
about
building
a
model
but
more
about
showing
what
it
takes
to
build
a
model
and
answer
production.
So
what
I
got
here
to
demonstrate?
That
is
a
very
simple
review,
app
so
say,
let's
say
you're
having
a
restaurant
and
then
you
want
to
highlight
the
positive
reviews
that
you
have
automatically
in
your
website.
So
I
go
and
ask
my
engineer
to
do
that.
B
What
do
they
do?
First,
they
build
a
very
simple
native
model
and
the
navy
model
is
as
simple
as
oh
I
know.
These
words
don't
match
like
being
negative
and
there's
these
other
words
that
are
positive,
I
mean
I.
Guess
it
works
a
bit,
but
it's
not
really
good
code
right.
So
how
can
you
improve
this?
Let's
use
machine
learning.
There
happens
to
be
really
good
NLP
models
and
tensor
a
tensor
happens
to
be
one
that
you
can
use
and
it's
like
really
prepackaged.
B
Well,
so
all
you
got
to
do
is
like
retrain
it
for
your
specific
use
case.
In
this
specific
use
case,
it's
going
to
be
how
do
I
get
reviews
to
be
digested
and
build
models
out
of
it.
So
let's
go
ahead
and
do
that.
What
I'm
going
to
do?
First
is
I'm
going
to
start
I
can't
type
and
speak
so
I'm
going
to
I.
B
Have
everything
scripted
here:
I'm
gonna,
first
start
training
locally,
so
I'm
here,
switching
to
my
local
mini
cube
cluster
and
the
general
idea
here
is
that
you
can
you
can
iterate
really
quickly
locally
on
your
workstation
or
on
your
laptop
then,
once
you
feel
comfortable
with
your
model,
then
you
can
take
it
to
the
cloud.
So,
let's
first
get
going
with
the
local
environment.
So
it's
this
is
all
I
did
for
literally
installing
cube
flow.
B
So
what
you
get
is
like
you
get
all
the
components
that
David
mentioned,
but
they
are
pre
packaged
in
a
way
that
it's
cubed
is
native
and
it's
really
easy
to
use.
So
you
don't
have
to
go
and
create
a
whole
bunch
of
objects
to
make
it
work
and
give
a
nice.
So
once
cube,
flows,
install
I
want
to
show
you
how
to
Peter
works
here,
for
example,
so
Jupiter
is
already
pre
packaged,
so
I
can
have
notebooks
locally
or
in
the
cloud.
B
And
since
you
get
super
hub,
you
can
choose
from
a
variety
of
pre-canned
images
or,
and
you
can
set,
you
can
change
the
resource
limits.
This
is
like
the
most
common
api
that
most
people
care
about,
and
it
also
hits
phone.
You
can
actually
have
a
notebook,
but
for
today
we're
not
going
to
use
notebooks,
but
we
are
rather
going
to
use
the
command
line
because
I
assume
we're
all
fans
of
cube
controllers
right,
so
I'm
going
to
switch
to
the
command
line.
So
what
are
you
here?
B
I
configured
my
model,
that's
packaged
through
case
in
it
I
configured
it
to
run
locally,
so
I
set
it
to
I,
set
the
PVC
name.
So
I
set
up
a
PUC
locally
so
that
I
can
use
the
data
that
is
available
locally
on
my
laptop
to
train,
rather
than
like
pulling
it
from
where
from
cloud
and
I
also
like
set
the
local
training
to
be
true.
When
I
did
that
and
then
I
went
ahead
and
deploy
the
model
locally.
B
B
B
Now
you
can
use
TP
use
from
any
of
the
nodes.
Pre-Pro
Shandra
TPU
4,
for
this
demo,
I'm
going
to
go
back
and
I
also
deployed
cube
flow
to
the
same
cluster
I'm,
going
to
change
the
configuration
to
make
it
work
in
the
cloud.
So
I
have
to
I
have
to
update
a
few
parameters
like
after
update
that
the
data
is
going
to
be
stored
on
Google,
Cloud
storage.
B
I'm,
also
going
to
say
that
my
output,
the
model
checkpoints,
have
to
be
stored
on
Google,
Cloud
storage,
because
you
have
many
different
parts.
Multiple
writers
and,
like
cloud
storage,
happens
to
be
the
best
interface
for
multiple
writers.
Once
I
did
that
I
went
ahead
and
deploy
the
model
and
then
now
I'm
also
going
to
deploy
a
model.
That's
using
TP
use
so
I'm
going
to
configure
my
TPU
model
and
then
go
ahead
and
deploy
that
against
the
same
GK
cluster.
B
So
now
that's
done
we're
gonna
wait
for
the
posture
on
and
could
use
the
GK
UI
here
to
figure
out
whether
you're
the
pots
that
we
deployed
are
running.
So
you
see
here
that
there's
a
TPU
pod,
that's
running!
That's
that's
running
the
TPU
training
model,
so
we
can
go
in
here
and
throw
the
logs
to
check
that
the
model
is
actually
running.
So
you
see
here
that
model
training
has
actually
started
and
if
you
go
back
and
if
you
look
at
the
GPU
model.
B
You'll
see
the
same,
that
training
is
actually
starting
and
it's
running
in
the
cloud.
So
if
you,
so,
you
probably
have
to
wait
for
a
bit
for
training
to
complete,
because
it's
going
to
take
like
almost
an
hour
for
the
model
to
be
fully
trained,
but
we
don't
have
that
time
here.
So
what
we're
going
to
do
instead
is
we're
going
to
use
a
prepackaged
model,
but
before
that,
I
also
wanted
to
show
you
that
you
get
tense
about
integration
out
of
blocks.
B
So
what
that,
if
you're
not
familiar
with
tenza
board,
it's
a
really
useful
UI,
that
that
gives
you
insights
into
how
your
training
is
actually
progressing
so
with
tons
about
prepackaged.
So
once
you
go
through
the
training,
what
you're
going
to
see
is
you're
going
to
see
graphs
like
this,
where
it
show
you
the
training
rate
like,
for
example,
here
the
training
read
what
GPUs
is
about
2.5.
If
you
switch
to
say
TP
use,
the
training
rate
is
going
to
be
like
much
higher
for
this
specific
model,
so
you
get
some
speed
boost
there.
B
B
Have
our
samples
about
the
restaurant
out
that
needs
to
be
optimized
I'm,
going
to
use
a
pre
pre-built
model
and
I've
deployed
that
model
to
the
same
GK
cluster.
So
let's
just
go
check
that
the
model
is
actually
running.
So
you
see
here
that
there's
a
deployment
here
which
is
actually
serving
the
bottle,
and
the
deployment
is
now
running
so
I'm
going
to
go
back
to
my
review,
app
right
and
earlier
what
we
saw
is
we
saw
that
we
saw
this
first
review
here
as
a
human.
B
We
know
that
it's
not
really
a
positive
review,
but
then
our
nave
classification
classified
it
as
a
positive
review.
Let's
see
if
the
machine
learning
model
does
any
better
okay,
there
you
go
so
it
was
able
to
generalize,
and
it
was
able
to
figure
out
that
there's
more
context
associated
with
language,
and
it
was
able
to
perceive
that
it's
actually
a
negative
sentiment
and
not
a
positive
sentiment.
B
So
this
is
to
show
that
you
have
a
lots
of
great
tools
and,
like
David,
said
it's
like
0.1,
so
that
there
will
be
things
that
you
have
to
fix
here
and
thereby,
then
you
get
something
that
gets
you
going
and
being
able
to
like
show
that
you
can
go
through
the
whole
journey
within
like
four
or
five
minutes
now,
and
so
you,
you
should
be
expecting
the
same
when
you
get
this
going.
I
hope
that
everyone
gives
it
a
try
and
feels
empowered
to
do
ml
with
that,
like
David.
A
So
you
know
I
I,
as
someone
who's
loved
ml
for
years,
I
was
absolutely
blown
away
for
when
I
saw
that,
for
the
first
time
the
amount
of
configurations
and
changes
and
installation
you
have
to
have
to
have
something
run
locally
and
then
run
in
one
environment
and
then
run
in
the
second
environment.
Attached
to
two
different
accelerators
today
is
crazy
and
you
saw
it
happen
on
stage
that
was
a
real
live
demo
of
it
actually
running
in
all
these
various
locations
and
experimenting
in
a
way
that
massively
increased
the
performance.
A
What
you
didn't
see
was
bespoke
solutions
every
one
of
those
packages.
We
were
using
encapsulate
an
open-source
ml
solution
today
that
with
large
communities
that
are
fully
production
ready,
and
we
are
really
really
excited
to
help
those
communities
come
use
cute
flow
as
a
way
to
get
their
solutions
out
to
the
market.
Second,
you
didn't
see
any
cloud
specific
tech.
Even
the
elements
of
running
on
GCP,
attaching
two
GPUs
attaching
GPUs
deploying
node
pools
those
were
abstracted
away.
We
used
the
exact
same
deployment
from
cube
flow
on
Prem,
as
we
did
on
the
cloud.
A
The
only
thing
we
did
was
change
some
parameters,
which
is
appropriate
because
you
want
to
point
it
at
the
place
where
it
is.
We
didn't
have
to
change
any
code
or
other
configurations,
and
you
also
didn't
see
forking
of
the
kubernetes
api
x'.
We
use
the
natural
extension
points
to
create
cube
flow,
CR,
d68,
tensorflow,
CR
DS,
that
include
master
servers,
parameter
servers
and
worker
nodes,
say
a
very
complicated
deployment
that
understands
how
to
communicate
with
each
other.
A
We
did
that
natively
and-
and
that's
really
where
we
are
today,
we
want
to
enable
this
brand-new
world.
In
this
case,
we
are
talking
about
data
science
and
data
science
will
impact
every
industry,
but
we
can't
ask
everyone
to
become
experts
in
everything
we
need
to
give
the
used.
Just
like
kubernetes
does
natural
extension
points
to
give
people
the
power
to
make
changes,
and
you
know
I
love
this
story.
A
They
were
plowing
the
streets
first,
instead
of
the
sidewalks,
and
the
sidewalks
happened
to
be
where
the
women
and
stay-at-home
parents
were
the
ones
who
walked
in
the
morning
and
so
Dubai
doing
nothing
else,
but
shifting
the
order
of
how
they
plowed
their
streets.
They
were
able
to
save
lives
in
priest
injury
hour.
Reduce
injury,
excuse
me
and
improve
people's
lives,
but
those
data
scientists.
They
are
to
me
those
people
operating
at
that
level.
The
public
policy
people,
the
snow,
plow
scheduling
people,
they
were
not
gonna,
understand
kubernetes,
Granada's,
api's,
anything
under
the
hood.
A
We
need
to
give
everyone
the
power
to
use
the
tools
that
we
are
building
and
we
need
to
do
that
by
providing
clean
extension
points
and
clear
ways
to
build
on
top
of
the
tools
that
we're
building
now
we're
just
getting
started.
We
have
over
70
different
contributors
to
the
platform
right
now
we
have
17
different
companies.
You
can
see
here
who
are
contributing.
A
You
have
great
use
cases
open,
faz,
released
a
really
cool
server
list,
solution
for
colorized
bot
against
cube
flow
and
we're
really
excited
and-
and
we
have
a
whole
bunch
coming
next
0.2
is
coming
very
shortly.
We
hope
this
summer
with
significantly
easier
setup,
better
abstraction
around
some
key
kubernetes
features,
auto
scaling,
accelerators
lots
more
packages,
I
mentioned
tensorflow
a
lot,
but
we
already
have
PI
torch
in.
A
We
have
a
cafe
proposal
and
we
really
do
want
every
single
ml
tool
to
be
deployable
via
cube
flow,
and
my
dream
really
is
that
in
call
it
a
few
years
time,
every
single
research
paper
in
the
world
has
three
things
associated
with
the
end
of
it:
first,
a
cube
flow
deployment,
second,
a
model
and
third,
a
data.
A
package
of
data
and
with
those
three
things
we
can
make
genuine
changes
and
improvements
around
the
reproducibility
problem,
but
most
of
all
cube
flow
is
open.
We
are
listening.
A
We
want
to
solve
real
problems
today
you
tell
us
how
we
are
open
in
all
the
ways.
It
is
possible
to
be
open,
github,
slack
twitter,
you
name
it,
but
we
would
love
more
than
anything
to
empower
all
of
you
to
take
this
and
make
kubernetes
the
platform
of
choice
for
ml
in
the
future.
Thank
you
very
much.