►
From YouTube: A Year of Democratizing ML With Kubernetes & Kubeflow - David Aronchick & Fei Xue, Google
Description
Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects
A Year of Democratizing ML With Kubernetes & Kubeflow - David Aronchick & Fei Xue, Google
A year ago, we introduced the Kubeflow project to make end-to-end ML pipelines on Kubernetes composable, portable & scalable. Today, thanks to passionate contributors from all over the world, we have the most popular ML platform for Kubernetes. At this Kubecon, we are announcing Kubeflow 1.0, graduating the project to generally available. In this talk, we will cover never before seen features: a web-based UI, simplified setup & sophisticated ML tooling including hyperparameter search and Google's TensorFlow Extended project. Additionally, we will be demonstrating the newly integrated Pipelines project wiring together multi-cloud ML with continuous training and hosted services. Thanks to Kubernetes native extensibility, we are able to bring ML to an entirely new audience, where as long as you can code, you can build complete end-to-end solutions.
To learn more click here: https://sched.co/FuLr
A
B
David
on
check,
I'm,
product
manager
and
co-founder
of
the
cube
flow
project,
so
I
guess
this:
is
it
okay?
So
with
that
you
know
this
is
a
really
awesome.
If
occasion
I'm.
This
is
the
first
obviously
cube
con
in
China.
This
is
my
first
talk
in
China
I've
been
representing
kubernetes
for
since
January
of
2015,
and-
and
this
is
really
really
I-
can't
tell
you
how
many
people
have
come
up
to
me,
asking
to
have
talks
in
China
about
kubernetes
and
cube
flow
and
ml,
and
so
I
hope.
B
B
You
know
so
many
people
have
so
many
different
takes
on
it.
Is
it's
something
just
incredibly
basic
like
advanced
linear
regression?
Is
it
something
that's
gonna
take
over
the
world
who
knows,
but
the
reality
is:
is
that
at
Google
and
across
the
machine
learning
industry
we
think
largely
it's
this.
It's
a
way
of
solving
problems
without
explicitly
knowing
how
to
create
a
solution
where
you
have
data
and
features
and
sources
for
so
many
different
things
and
your
look.
B
What
you're
really
looking
to
do
is
get
to
an
answer
yet
you're
not
exactly
sure
how
two
things
correlate
and
that's
a
really
powerful
thing,
and
we
went
into
cute
flow
thinking
this
very
thing
at
Google
everything
at
Google.
It
is
touched
by
AI
in
some
way,
and
one
of
the
most
profound
ways
is
this
example
that
we
have
here
with
Google
data
center
operations.
At
you
know,
obviously,
data
centers
are
one
of
our
most
expensive,
most
costly,
most
important
things
that
we
do
and
for
us
we
hire
great
engineers.
B
We
think
about
problems
very,
very
deep
when
we
took
at
that
knowledge,
when
we
took
all
the
data
that
we
collected
and
we
applied
machine
learning
to
it.
We
found
out
a
really
amazing
thing
that,
even
though
these
people
were
great,
they
were
the
best
data
center
engineers.
Arguably,
in
the
world,
we
were
able
to
make
huge
progress
against
them
and
you
can
see
it
here.
This
is
a
way
to
measure
data
center
effectiveness.
B
It's
called
power
usage
effectiveness
where
you
measure
the
amount
of
power
going
in
and
and
the
amount
of
computation
that
you
get
out
the
other
side.
Ideally,
you
want
to
be
as
low
as
possible,
meaning
you
have
the
lowest
possible
power
going
in
and
you're
getting
the
most
power
powerful
computation
out
of
it.
In
this
case
it
started
out
very
high.
B
B
They
want
to
just
think
about
models
and
pipelines,
and
things
like
that
and
the
reality
is,
is
that
you
know
they
had
to
think
about
it,
and
so
what
we
wanted
to
do
was
think
about
changing
the
game
with
machine
learning
in
the
same
way
that
containers
and
kubernetes
changed
it
for
overall
deployment
and
create
something
called
cloud
native
apps.
Now
the
question
is:
could
we
get
to
cloud
native
ml?
B
What
would
it
even
look
like
for
us?
It
looks
kind
of
like
the
following
when
people
think
about
machine
learning
today,
they
really
think
about
a
single
element
right.
They
think
about
building
a
model.
Are
you
gonna
use?
Paddle
paddle
or
tensorflow
or
psyche
it-
you
know,
what's
the,
how
quickly
does
it
converge?
What's
the
accuracy
level
and
so
on?
B
But
the
reality
is
is
that
a
machine
learning
pipeline
looks
a
lot
more
like
this,
where
you
have
lots
of
micro-services
loosely
coupled
together
and
what
you're
trying
to
do
is
get
to
a
solution
that
that
involves
actual
production
readiness,
even
if
you're
a
data
scientist
working
at
the
beginning,
and-
and
this
is
what
that
data
scientist
experience
looks
like,
whereas
the
first
one
might
be
relative
to
the
platform
and
the
various
services.
A
data
scientist
wants
to
do
something
completely
separately.
They
want
to
just
be
at
Jupiter
level
or
docker.
B
They
want
to
train
their
model,
they
want
to
deploy
it
and
they
want
to
integrate
it
into
their
app
as
quickly
as
they
can
and
and
worse
than
that,
even
with
that
complex
pipeline
I
haven't
even
started
to
get
to
that
where
the
real
problems
are,
which
is
around
the
entire
stack
end
to
end.
You
start
at
the
model
and
you
walk
all
the
way
down.
Webinars
you
axe,
tooling,
storage,
runtimes.
B
These
are
all
things
that
you're
gonna
have
to
deal
with
and
in
that
entire
pipeline,
that
I
was
talking
about
that's
just
the
very
top
of
it
all
the
other
things
kind
of
sit
under
the
hood
and
now
to
make
it
even
more
complicated.
It
doesn't
just
sit
into
the
hood
in
one
environment.
It
sends
over
the
hood
in
multi
cloud
environments,
where
you
not
have
to
replicate
that
entire
thing
over
and
over
again
in
this
case,
81%
of
enterprises
have
more
than
one
cloud
and
they
don't
just
have
more
than
one.
B
They
have
up
to
five,
which
makes
your
problem
much
much
more
complicated,
whereas
this
might
be
your
experimentation
one.
You
go
one
step
further
and
have
to
replicate
that
it's
same
environment
for
training
and
then
finally,
your
same
environment
for
cloud.
So
this
is
a
very,
very
complicated
problem,
and
this
right
here
is
why
we
introduced
cube
flow
in
December
of
2017
at
cube
con
Austin.
We
introduced
this
portable
framework
that
let
you
declaratively
describe
a
very
complicated
machine
learning
framework
and
then
pipeline.
B
What
using
the
elements
of
kubernetes
to
make
your
deployment
and
management
much
easier.
Our
mission
remains
the
same
a
year
in
we
want
to
make
it
easy
for
everyone
to
develop,
deploy
and
manage
portable
distributed
ml
on
kubernetes,
and
what
that
looks
like
is
this:
you
have
that
base
layer
kubernetes.
You
know
that
it
runs
the
same
everywhere
as
dan
cone
talked
about
earlier.
The
investments
that
the
CN,
CF
and
kubernetes
projects
make
in
conformance
means
that
you
know
no
matter
where
you
go.
B
You're
gonna
be
able
to
plug
into
a
rich
kubernetes
deployment
that
has
a
minimum
bar
for
the
api's
and
other
objects
available
and
then
on.
The
left
you
describe
your
pipeline
using
cube
flow
and
cube
flow
is
something
that
you
can
stamp
out
in
each
one
of
those
locations
in
a
in
a
declarative
way.
Knowing
that,
because
you
have
kubernetes
to
base
it
on
you're
gonna
be
able
to
replicate
that
very
quickly
and
what
that
means
is.
We
now
do
have
cloud
native
ml.
B
We
have
a
portable
framework
which
operates
at
the
cloud
level
loosely
coupled
together
micro,
service-oriented
in
the
years
since
we've
launched.
We
have
a
tremendous
amount
of
momentum.
1900
commits
over
a
hundred
community
contributors
with
30
different
community
companies,
contributing
including
the
list
you
see
here
and
many
more
and
when
I
say
what
it
means
to
have
a
rich
community.
It
looks
like
the
following
when
you
think
about
one
of
the
most
successful
communities
in
the
world
kubernetes,
it
looks
something
like
this
and
I
know.
B
That's
a
lot
of
different
companies
contributing
in
a
lot
of
different
ways.
Let
me
kind
of
simplify
it
for
you.
It
looks
like
this.
That's
Google,
and
this
is
not
Google
and
that's
really
important
right.
We
are
not.
We
at
Google
are
not
the
number
one
contributors.
Everyone
else
is
that
is
so
important
and
it
took
us
years
to
get
here.
You
know
the
first
year
second
year
it
we
were
the
over
50%.
B
That
was
incredibly
important
for
us
to
get
to
that
point
where
other
people
could
find
the
place
in
the
ecosystem
to
make
it
then
around
4q
flow.
It
looks
like
this.
Here's
two
different
companies
contributing
and
to
simplify
it
again,
Google
and
not
move
and
we're
already
over
50%
of
our
contributions
come
from
non
Google.
This
is
what
healthy
community
looks
like
and
let
me
take
it
one
step
further.
They
aren't
folks
just
aren't
working
on
bugs
and
I'm,
not
saying
bugs
aren't
incredibly
important.
They
are.
B
We
know
we
want
to
get
to
stability,
but
the
reality
is.
It's
really
significant
contributions
that
move
our
platform
forward
very
very
quickly.
For
example,
we
have
a
great
framework
in
the
core
of
the
platform
called
cat
tube.
It
helps
do
hyper
parameter,
optimization,
it's
plugged
in
and
in
particular
something
PHA
will
show
in
just
a
minute.
We
make
it
very
very
easy
to
do
this
even
with
very
little
code
and
that
came
from
NTT
and
we
couldn't
have
been
happier
about
it.
B
Second,
we
have
a
brand
new
serving
framework
which
literally
just
launched
two
months
ago.
I
believe
it
is
at
this
point
from
Nvidia,
it's
called
tensor
RT
and
it
makes
serving
on
GPUs
much
much
more
efficient,
including
multiple
models
per
GPU
per
node,
heterogeneous,
GPUs
and
and
environments,
and
it
integrates
with
the
underlying
auto
scalars
and
latency
and
Health
metrics,
again
external
company,
contributing
to
the
core
of
cube
flow.
This
is
incredibly
important.
One
more
example
is
our
go
from
Intuit
I
mentioned
that
our
go
is
already
in
the
box.
B
This
is
an
incredibly
powerful
system
for
workflow
and
and
orchestration
and
pipelines
we'll
touch
on
that,
and
it
really
allows
for
doing
get
ops
you
might
have
heard
for
launching
it.
Excuse
me
for
it
for
declaratively
rolling
out
your
machine
learning
pipeline.
So
that's
everything
now
I'd
like
to
hand
off
the
fade
to
talk
about
cue
flow
0.3,
which
we're
happy
to
announce.
A
So
we
recently
really
use
cube
flow,
0
point
3
and
the
zero
point
3.
There
are
two
major
seams
that
we
really
invested
a
lot
in
so
once
around
how
to
make
you
flow
easier
to
deploy.
So
you
can
have
a
much
better
getting
to
started
experience.
The
other
part
we
want
to
make
investment
in
was
how
to
make
the
development
experience
on
cue
flow,
more
powerful
and
able
to
use
in
different
use
cases.
A
So,
on
the
deployment
side,
we
released
a
new
deployment,
CLI
called
a
KF
controlled
sh,
it's
a
bash
script
where
you
can
use
to
install
all
the
Q
full
components
just
from
one
command.
I'm
gonna
show
more
of
that
later
as
well.
The
second
thing
for
deployment
is
that
we're
having
a
much
much
simpler,
set
up
and
later
I'm
going
to
show
a
one-click
to
deploy
of
how
you
can
provision
the
infrastructure
as
well
as
install
all
the
Q
point.
Deployment
was
just
one
simple:
click
on
the
deploy
one
development
side.
A
We
are
also
making
a
lot
of
different
investment
and
I
won't
just
want
to
highlight
two
of
them.
One
is
our
PI
torch
operator.
When
we
first
got
started,
we
focused
on
TF
job
and
now
we're
look,
we're
listening
to
the
community
and
we're
hearing
that
people
care
about
using
PI
torch
on
cue
flow,
and
now
our
PI
torch
operator
has
feature
parity
with
TF
job,
while
the
biggest
obstacles
to
getting
quality
models
is
to
make
sure
that
you
have
the
proper
hyper
parameters
for
your
models
and
that's
why
hyper
parameter
tuning.
A
It's
super
super
important
to
us
and,
as
part
of
cue
flow
0.3,
we
announced
a
hyper
parameter
tuning
CRD.
It's
called
study
job
that
enables
you
to
do
hyper,
parameter
tuning
and
searching
without
needing
to
write
any
code.
So
I
talked
about
the
general
themes
of
what
we
have
in
kill
queue
flow
0.3.
Let's
double
click
on
some
of
the
things
and
I'll
show
a
couple
demos,
so
the
first
one
is
for
deploy.
We
have
something
called
quick
to
deploy.
A
We've
been
hearing
from
the
community
that
it's
very
difficult
to
get
started
with
queue
flow
because
you
need
to
install
packaging
libraries
and
then
you
need
to
download
and
install
the
different
components
that
comes
with
SKU
flow.
We
want
to
make
that
getting
started
experience
much
much
easier.
So
what
we
came
up
with
is
a
one-click
installation
tool.
A
It
is
available
to
you
with
a
very
clean
web
interface,
so
it's
a
web
app
and
the
way
you
would
use
it
is
basically
it
uses
a
bootstrap
container
and
there's
this
one
bash
script
that
you
can
use
and
it
has
all
the
necessary
dependencies
included-
was
a
bash
script.
You
can
then
use
it
in
any
infrastructure
deployment
tool,
so
that
would
be
deployment
manager
on
GCP
or
CloudFormation
IDs.
And
finally,
what
you're
seeing
we're
gonna
see
now
has
Note
happening
to
need
it
at
all.
A
A
All
you
need
to
do
is
specify
what
project
you
want
to
deploy
in
the
name
of
the
deployment
and
some
security
configurations,
then
the
click
to
deploy
it
will
actually
provision
the
infrastructure
on
top
of
GCP,
meaning
that
there
will
actually
be
a
cluster.
The
spinning
up
that
will
host
your
cube
flow
components.
So,
as
you
can
see,
after
I
did
the
click
to
deploy
now
I
have
AQ
flow
cluster,
it's
provisioned
through
the
click
to
deploy
on
GCP.
A
A
A
The
second
thing
that
we
want
to
highlight
is
Q
plop,
I
clients.
This
was
announced
very
recently
and
what
we
see
and
noticed
the
community
is
that
machine
learning
solutions
are
often
multi-staged.
There
are
a
lot
of
components
currently
in
Q
flow
that
deals
with
each
of
the
stage,
but
it
was
very
difficult
to
pull
all
the
different
stage
and
components
together
and
that's
why
we
have
to
flow
pipelines.
It
is
a
micro
service
platform.
It
is
designed
to
enable
reusable
components
and
workflow
orchestration.
A
So
it
is
a
one
place
where
you
can,
orchestrate
all
the
different
components
and
run
machine
learning
and
to
end
what
it
is
is
that
is
a
Python
SDK
and
you
can
use
the
Python
SDK
to
describe
and
containerize
your
machine
learning
tasks.
We
choose
Python,
because
this
is
very
natural
to
data
scientists
so
that
they
can
read.
A
We
can
really
meet
where
they
are
so
that
they
can
themselves
describe
what
they're
trying
to
do
with
their
whole
machine
learning
task
after
they
describe
their
pipeline
in
the
Python
SDK,
we
will
compile
it
to
argl,
which
is
coming
already
in
the
box
from
Q
flow,
and
we
also
as
pipelines.
We
offer
a
front-end
GUI
to
do.
Experiment
logging,
as
well
as
analytics,
what's
really
important
about
this,
is
that
all
of
these
steps
are
containerized.
A
A
So
in
this
pipeline
we
will
use
a
very
simple
training
example
that
utilizes
GPUs,
so
in
this
Python
script,
as
you
can
see,
we
have
two
main
operators.
The
first
operator
is
the
training
operator,
and
the
second
operator
is
the
post-processing
operator
in
the
training
operator,
because
this
is
all
container
base
we're
getting
a
image.
This
is
where
our
images
is
a
very
simple,
straightforward,
Minnis
example
and
we
are
using
GPU
to
Train
it
as
each
step
of
pipelines.
A
It
has
a
series
of
the
input
parameters
as
well
as
output,
so,
for
example,
in
this
training
container,
it
has
three
inputs
that
we're
giving
it
so
the
first
one
is
learning
rate.
Second,
one
is
number
of
layers,
so
basically
how
many
layers
you
have
in
your
neural
network
and
the
third
one
is
what
kind
of
optimizer
you're
choosing
so
that
constructs
a
single
operator.
The
second
operator
is
supposed
processing
is
just
like
vanilla,
post-processing
step,
and
here
is
where
we're
actually
chaining
those
two
together.
A
A
So
the
second
Sep
I
will
actually
compile
my
pie,
so
script
here
is
the
yellow
file
that
I
generated.
So
as
I
mentioned,
the
whole
point
is
that
we
want
to
meet
data
scientists
where
they
are,
and
in
this
case
is
actually
Python
and
from
the
Python
we
generate
the
yellow
file
and
that
can
be
consumed
actually
by
cube
flow
and
kubernetes.
A
As
you
can
see,
the
yellow
file
generated
is
quite
long,
so
here
what
you
see
here
is
the
actual
pipeline
GUI.
It
comes
out
of
the
box
today,
with
cube
flow
on
the
left
side,
you
can
see
that
we
offer
pipelines
experiments
and
on
the
right
side,
it
comes
out
of
the
box
with
a
series
of
sample
pipelines
that
you
can
explore
into
so.
Let's
click
into
one
of
those
sample
pipelines
and
look
at
what
it
looks
like
so
in
this
sample
file.
A
A
Okay,
after
we
saw
the
sample
steps,
let's
take
what
we
just
saw
the
Python
file
and
the
llamo
they
compiled.
We
can
upload
the
Yamaha
into
the
pipeline's
UI,
and
now
you
have
a
new
pipeline.
That's
on
the
you
on
the
GUI,
so
we
have
the
concepts
of
experiment
and
runs
because
the
goal
of
the
pipelines
is
to
make
sure
that
you,
you
can
reuse
your
component
as
much
as
possible.
So
each
experiment
can
have
multiple
runs,
each
with
different
input
parameters.
A
A
A
Let's
go
to
the
next
components
that
we
want
to
highlight.
So
this
is
we
want
to
talk
about
auto
scaling
and
the
reason
we
want
to
talk
about
it
is
that
today,
if
you
want
a
data,
scientist
in
the
organization
wants
to
have
something
once
after
they
have
a
model
working
and
when
they
want
to
train
things
in
scale.
They
have
to
talk
to
somebody
else
in
the
organization,
the
IT
ops
person,
so
our
data
scientist
say
my
motto
is
working
great,
but
I
need
six
nodes.
A
The
IT
ops
person
is
there
to
help,
but
she
is
really
pretty
busy,
so
it's
gonna
take
some
time
for
this
to
actually
happen.
In
the
mean
time,
data
scientist
kind
of
is
welcome.
What's
going
on
so
when
the
IT
ops
person,
yes
to
it,
she
provisioned
six
note
the
data
scientist
is
happy
and
continues
to
do
her
work
and
after
doing
lots
of
work,
all
the
training
is
done.
She
is
kind
of
wanting
to
move
on
to
other
things,
to
finish
her
job.
A
So
that's
why
we
have
integrated
jobs
with
auto-scaling
feature.
So
our
goal
is
that
the
data
scientist
just
needs
to
describe
the
job.
Crewmen
ATIS
will
take
care
of
the
rest,
so
that
that
includes
the
CPUs
are
looking
for
the
memory
as
well
as
any
accelerator.
That's
needed
so
be
a
GPU
CPU
FPGA
TPU
what-have-you.
When
the
job
is
done,
the
resources
themself
are
going
to
be
automatically
deleted,
so
things
will
scale
down
by
themselves
and
these
resources.
A
A
The
GPUs
are
provisioned
automatically
with
cue
flow
and
kubernetes.
When
the
job
is
done,
the
resource
are
gone
automatically,
so
they
don't
need
to
worry
about
over
provisioning
or
overspending
or
resources
sitting
idle
and
our
IT
ops
people.
She
will
be
happy
because
now
she
has
free
cycles
to
do
things
that
matter
or
it
may
be
talked
about
YouTube.
A
A
So
what's
going
to
happen
there,
when
we
look
at
the
workloads
we
see,
our
pipeline
is
currently
initially
unscheduled
and
when
we
go
there
we
can
actually
see
the
reason
behind
it
is
that
there
is
not
enough
GPUs,
that's
required
to
do
the
training,
so
there
is
a
feature
called
knowled
ultra
provisioning.
That's
able
to
detect
the
reason
why
the
drop
is
not
being
able
to
scheduled,
and
now
it
is
working
to
spin
up
a
new
node
pool.
That's
able
to
accommodate
your
job.
A
A
A
Earlier
we
talked
about
how
hyper
parameter.
Chaining
is
super
important
to
how
data
scientists
do
their
work
and
now
I'm
going
to
show
you
how
this
will
work
in
the
Q
blow
scenario,
so
in
Q
flow
0.3,
we
a
now
study
job,
which
is
the
CR
D
for
hyper
parameter
tuning.
What
you
can
do
here
is
that
you
want
to
have
a
piper
parameter
tuning
job,
to
make
sure
that
we
have
the
right
input
parameters
for
your
training
job,
and
here
I've
applied
a
yellow
file.
A
You
can
see
that
this
is
our
study,
job
spinning
up,
and
we
could
have
a
quick
look
at
what
that
study.
Job
looks
like
so
this
would
be
how
I
describe
my
hyper
parameter
tuning.
In
this
case,
you
can
see
that
I'm
actually
trying
to
tune
three
different
things.
The
first
thing
I'm
trying
to
tune
is
the
learning
rate.
The
second
thing
is
the
number
of
new
or
network
layers,
and
the
third
thing
is
what
kind
of
optimizer
I'm
looking
for.
A
A
A
As
part
of
catted,
our
hyper
parameter
chaining
component,
it
comes
with
its
own
component,
called
mono
dB.
So
this
is
a
front-end
UI.
That
shows
it
will
help
you
visualize
and
analyze.
The
results
of
your
hyper
parameter
tuning
jobs.
So
in
this
example,
you
can
see
on
top.
So
the
y
axis
is
a
validation
accuracy,
and
each
of
these
dots
represent
a
new
run
of
job.
We're
trying
to
see
so.
You
can
see
that
some
of
the
runs
are
have
very
low
accuracy
around
10%
and
some
of
them
have
higher
accuracy
around
99%.
A
A
So
the
first
thing
we
try
to
do
is
we
try
to
do
a
group
buy
on
the
different
optimizers
and
we
want
to
see
their
average
validation
accuracy.
So,
as
I
mentioned,
we
have
we're
looking
at
three
different
optimizers
and
from
this
graph
we
could
tell
immediately
that
one
optimizer
was
trouble,
so
we
could
either
select
the
Adam
operator
or
the
SGD
operator
and
furthermore,
we
want
to
do
a
little
more
drill
down.
So
let's
say
we
only
filter
on
the
runs
that
uses
SGD
as
the
optimizer
and
then
from
the
model.
A
Db
UI
you'll
be
able
to
group
by
on
the
different
metric.
So
this
time,
let's
say
group
by
a
number
of
layers-
and
in
this
case
you
can
see
either
it's
either
two
layers
or
three
layers
doesn't
really
make
a
difference,
your
accuracy,
so
that
doesn't
matter
and
then
have
a
look
at
learning
rate
and
similar
things
apply.
So
after
we
have
the
insight
into
our
hyper
parameters,
we
can
go
into
our
pipeline
and
then
do
a
second
run.
This
time
we
know
what
to
put
in
for
the
optimizers.
A
A
And,
what's
most
is
that
you
can
see
now
we
used
to
have
six
notes
to
note
pools
with
three
different
GPUs,
but
after
we
run
our
two
pipelines
as
well
as
a
hyper
prompt,
there's
opportunity,
job
we're
back
to
three
notes.
This
is
exactly
where
we
started.
Instead
of
having
to
manually,
tune
down
or
spin
down
your
nose.
That
has
a
lot
of
resources
hash
to
it.
We
ended
up
spinning
it
down
automatically
for
you,
so
you
don't
really
have
to
worry
about
any
of
that.
The
auto
scaling
really
comes
to
the
rescue.
B
So,
thank
you
so
much
Faye.
We
are
just
getting
started.
We
are
quickly
marching
towards
1.0
and
would
love
to
hear
your
feedback.
What
are
the
things
that
are
missing?
What
are
critical,
particularly
around
enterprise
requirements,
things
like
upgrades
integration
with
I
am,
and
our
back.
One
of
the
things
we've
heard
a
lot
about
is
how
to
be
more
closely
tied
with
Jupiter.
B
We
know
that
Jupiter
is
the
place
where
data
scientists
are
working
right
now,
and
we
want
to
make
it
very,
very
easy
to
tie
those
together
and
provide
a
much
richer
pipeline
experimentation
and
model
management.
These
are
core
things
that
we've
already
heard
about
we're
trying
to
get
to
them
by
1.0,
but
we'd
really
love
to
hear
what
you
would
like
next
and
I
really
want
to
stress
this.
It
is
a
brand-new
world.
B
The
people
in
this
room,
the
people
watching
this
on
the
video
you
have
the
ability
to
make
huge
changes
in
industry
I
strongly
ask
you
all
to
think
deeply
about
ethical,
sustainable
ml
that
can
help
the
world
move
forward,
because
we
really
do
have
a
chance
to
make
an
enormous
change
and
things.
We
hope
that
things
like
cube
flow
will
help
enable
that
and
give
the
power
to
two
data
scientists
to
make
that
happen
quickly.
B
We
are
open
open.
You
know
in
every
way,
it's
possible
to
be
open.
Community
design,
open
source
come
to
github.
This
is
not,
by
any
stretch,
a
Google
only
project.
We
want
to
continually
meet
your
needs,
so
you
don't
have
to
go
out
and
build
your
own
bespoke
platforms
and
with
that
off
we
go
any
questions.
Yeah.
B
B
B
So
the
question
was:
if
I
understood
during
the
hyper
parameter
tuning,
we
store
the
outputs
of
the
various
choices
that
the
hyper
parameter.
Optimizer
was
using.
We
don't
store
the
artifacts
of
the
model.
We're
just
doing
a
cat
tib
is,
is
storing
the
the
metadata
related
to
the
model
in
model
DB
and
that's
what's
allowing
the
comparison
in
order
to
do
a
full
model
run,
you
would
then
do
exactly
I
faded
on
the
board.
You
would
take
that
metadata
put
it
back
into
your
pipeline
and
then
that
would
do
a
full
training
run.
B
B
You
yes,
so
the
question
was:
if
you
are
new
to
machine
learning,
should
you
use
cube
flow
to
get
started,
and
is
it
feasible
to
use
it
on
a
single
computer?
The
direct
answer
is,
you
can
certainly
experiment
Jupiter
and
Conda,
and
these
things
are
great
and
there
are
great
ways
to
try
out
machine
learning
and
we
recommend
it.
B
That
said,
the
purpose
the
design
behind
cube
flow
is
to
work
great
all
across
the
spectrum,
whether
or
not
you're,
on
a
single
laptop
like
this
all
the
way
up
to
a
must
multi
cluster
deployment,
and
the
benefit
is
because
you're
describing
your
entire
machine
learning
pipeline
using
very
standard
tools
that
run
great
anywhere,
including
mini
cube
on
your
laptop.
It
means
there's
very
little
change
when
you
decide
to
graduate
to
larger
and
larger
machine
sizes.
B
It'll
work,
fine,
no
matter
where
it
is
now
to
be
clear
on
your
local
laptop
you're,
not
gonna,
want
to
download
a
petabyte
of
data,
and
you
know:
do
a
full
training
run.
It's
great
for
experimentation.
Jupiter
comes
in
the
box,
tensorflow
job
comes
in
the
box,
pi
torch
comes
in
the
box,
--cat
tube
comes
in
the
box.
These
are
all
things
that
are
very
trivial
or
they
may
not
all
come
into
the
box
very
trivial
to
install
and
you
can
get
it
running
even
locally
and
see
all
the
richness
that
you
saw
here.
B
It
is
explicitly
to
be
clear
with
everyone.
It
is
explicitly
our
goal
to
be
as
easy
or
easier
than
just
going
and
doing.
Pip
install
foo
to
get
starting
with
machine
learning,
and
we
know
we're
not
there
yet,
but
with
the
investments
that
we
made
around,
you
know
KF
control
and
other
things
like
that.
We're
hoping
to
get
there
very
quickly.