►
From YouTube: Machine Learning as Code: and Kubernetes with Kubeflow - Jason " Jay" Smith & David Aronchick
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
Machine Learning as Code: and Kubernetes with Kubeflow - Jason " Jay" Smith, Google & David Aronchick
Machine Learning has become an increasingly popular topic in the world of data. At the same time, the concept of microservices through containerization has increased in popularity as it allows for developers to create and package applications for easy export and distribution through various clouds.Kubeflow is an open source project lead by Google to merge both concepts, allowing users to leverage the power of Kubernetes to run the training and serving of their ML models.This convergence of technologies does result in a new way to think of Machine Learning. We now think of can think of machine learning as code bundles. My session will show how with Kubeflow and GitOps tools, you can go beyond simply deploying and training TensorFlow models but bundling the entire infrastructure into a code package and treat the entire machine learning process as a pipeline.
To Learn More: https://sched.co/GrVh
A
Hey
everyone
welcome
to
a
machine
learning
as
code
a
year
of
democratizing
ml
with
kubernetes
NQ
flow,
if
you're
not
here,
for
machine
learning
or
kubernetes
or
cube
con
or
keep
flow.
Well,
I,
don't
know
what
you
did
in
this
room,
but
you're
gonna
get
a
whole
hour
of
it
or
40
minutes.
So
one
year
ago,
two
days
ago
we
introduced
cube
flow
and-
and
we
really
started
with
the
question
that
you
see
here
before
you-
machine
learning-
is
everywhere.
Everyone
is
talking
about
it.
It's
solving
so
many
solutions
today,
but
it's
this
vague
concept.
A
Guy
I
do
a
bunch
of
math
I
backport
some
weights
and
magically
all
of
a
sudden
I
get
answers,
and
that's
really
strange
for
so
many
people,
not
just
because
it's
this
new
advanced
concept,
but
because
they
didn't
know
how
to
go
about
even
doing
the
first
step
of
implementing
a
machine
learning
solution.
So
at
Google
we
had
this
problem
and
we
came
up
with
a
definition
and
the
definition
was
machine.
Learning
is
a
way
of
solving
problems
without
explicitly
knowing
how
to
create
the
solution.
A
When
you
take
an
enormous
amount
of
data
and
you
label
it
and
you
think
about
it,
deeply
and
you're
not
really
able
to
figure
out
exactly
what's
going
on,
that
might
be
a
space
where
machine
learning
can
can
go
through
and
process
all
this
stuff
and
come
up
with
a
better
answer
for
you
and
and
at
Google.
We
took
this
and
it
changed
our
entire
company.
The
number
of
people,
the
number
of
groups
and
divisions
who
are
using
machine
learning
today
is
enormous,
and
this
isn't
just
kind
of
like
toy
projects
or
next-gen
projects.
A
This
is
deeply
wired
into
the
way
that
we
decided
to
start
changing
our
business.
In
this
case,
I
like
to
give
this
example
around
data
centers,
so
at
Google.
Obviously,
data
centers
are
one
of
the
largest
expenses
networking
and
electricity
and
cooling.
These
are
the
things
that
run
every
single
query:
every
single
YouTube,
video,
every
single
Android
query,
every
single
Gmail
query
costs
in
a
norm,
amount
of
money,
and
we
hire
some
of
the
best
engineers
for
to
solve
these
problems.
A
You
know
they
know
more
about
datacenter
cooling
than
than
anyone
in
the
world.
Arguably,
and
and
yet
we
still
were
spending
a
lot
of
money
on
it
and
the
reality
is,
is
that
it
we
looked
at
it?
We
were
collecting
so
much
data.
We
thought.
Could
machine
learning
help
solve
this
problem
for
us?
What
they
did
was
they?
A
They
took
all
the
data
and
they
collected
it
and
they
wired
it
up
to
a
bunch
of
different
algorithms
and
pipelines
and
so
on
and
they
went
and
they
tried
to
see
if
they
could
improve
things,
and
so
here
they
they
started.
Measuring
and
the
measurement
for
data
center
efficiency
is
power,
usage
effectiveness,
the
amount
of
power
that
goes
in
and
the
number
of
computational
cycles
that
you're
actually
able
to
achieve
they
went
in.
A
They
would
turn
on
these
machine
learning
models
and
almost
instantly
the
cost
for
the
machine,
learning
or
the
cost
for
those
data.
Centers
went
down
and
the
instant
they
turned
it
off.
They
went
back
up
and
that's
how
dramatic
it
was
again,
the
best
data
center
engineers
in
the
world,
thinking
about
renewable
thinking
about
efficiency,
thinking
about
cost
savings.
They
got
beat
almost
instantly
by
these
machine
learning,
algorithms
and
what
they
found
was
in
many
cases
it
wasn't
even
decisions
that
they
weren't
sure
how
to
do.
It
was
stuff.
A
It
was
many
many
thousands
of
small
decisions
that
these
engineers
were
like.
Well,
that
can't
possibly
add
up,
but
the
machine
learning
system
understood
that
it
in
fact
could
add
up
and
did
make
a
huge,
huge
difference,
but
the
reality
is
that
machine
learning
is
hard
and
it's
not
just
hard
because
I'm
bad
at
math
it's
hard,
because
there
are
so
many
components
that
you
have
to
think
about
when
it
comes
to
machine
learning.
How
do
you
roll
out
containers?
How
do
you
roll
out
new
systems?
How
do
you
make
upgrades?
A
How
do
you
build
pipelines
to
connect
all
these
systems
together
and
what
we
decided
when
we
started
cute
flow
was
to
think
about
containers
and
kubernetes
as
the
way
to
go
about
it
and,
and
the
reason
we
thought
about
that
is
because
of
a
big
movement
called
cloud
native
applications.
How
do
you
break
your
complex
monolithic
application,
down
into
a
series
of
small
components
that
are
loosely
coupled
together,
that
understand
how
to
scale
up
how
to
move,
how
to
be
portable,
but
between
environments?
And
so
we
thought?
A
Could
we
introduce
cloud
native
ml
and
cloud
native
ml
looks
like
the
following
when
you
think
about
ml
today,
most
people
or
many
people
will
say:
well,
it's
all
about
building
model
and
I'm
gonna
choose
PI,
torch
or
chain,
er
or
tensorflow.
What
what
is
am
I
going
to
choose,
but
the
reality
is
it's
not
that
it's
all
of
these
components
together?
It's?
How
do
you
get
your
data
in?
How
do
you
transform
it
validated?
Do
a
distributed?
Training
run,
build
it,
deploy
it
roll
it
out
to
production,
monitor
ant
login.
A
These
are
all
independent
micro
services
that
many
times
you've
already
started
to
address
via
your
move
to
cloud
native
applications,
and
it's
not
just
that.
That's
on
the
just
the
infrastructure
side,
when
you
think
about
the
end
user
side
deploying
and
queue
flow,
using
Jupiter
using
docker
images,
training
at
scale
all
these
things.
Those
are
additionally
areas
where
data
scientists
really
don't
want
to
think
about
all
these
other
components.
They
really
want
to
think
about
the
underlying
effort.
A
The
solution
they're
trying
to
come
up
with,
and
not
really
worry
about
this
stuff
under
the
hood
and
when
I
say
a
complex
stack
of
micro
micro
services,
it's
not
just
that
top
layer
that
I
was
pointing
out
there.
It's
everything
under
the
hood,
it's
drivers
os's.
How
do
you
do
orchestration
between
nodes
as
they
need
upgrades
so
on
and
so
forth?
And
unfortunately
it's
not
just
about
one
cloud
either.
The
reality
is
is
that
81
percent
of
enterprises
use
multiple
clouds
today
and
they
don't
just
use
multiple
clouds
or
two
clouds.
A
They
use
up
to
five.
So
you're.
Now
talking
about
a
situation
where
you
have
your
experimentation
and
then
you
didn't
need
to
replicate
that
exact
same
stack
in
a
new
place
when
you
do
training
and
then,
finally,
when
you
potentially
move
to
the
cloud
or
some
highly
scalable
environment,
and
that's
all
the
motivation
behind
why
we
decided
to
launch
cube
flow.
That's
me
a
year
ago,
slightly
less
spray
here,
maybe
with
vish
the
other
co-founder
and
then
Jeremy
lui,
who
I
saw
on
the
audience
somewhere.
A
So
that's
what
we
say
when
we
say
cloud
native
ml.
In
the
past
year
we've
had
tremendous
momentum.
Over
1,900
commits
hundreds
of
community
contributors,
30
plus
can
companies
contributing,
including
a
set
of
the
people
that
you
see
there
and
though
it
certainly
started
at
Google.
It
is
not
remotely
Google
only
github
LinkedIn
Intel,
the
Jupiter
project,
Red
Hat
celled
in
IBM.
These
are
just
a
small
set
of
people
who
have
already
contributed
upstream
and
helped
make
cube
flow,
a
truly
community
driven
project,
and
by
that
I
mean
the
following.
A
The
reality
is
is
when
you
think
about
a
very,
very
successful
community
it'd
be
hard
to
do
better
than
kubernetes
and
when
Cooper
days
first
came
out
and
I
know
because
I
was
there,
people
would
like.
Oh,
this
is
Google
only
you
know,
you're
gonna,
somehow
tweak
it
in
some
way.
That
makes
it
only
work
on
Google
cloud
and
I
would
go
out
and
I
would
say
many
many
times
nope,
not
us.
We
want
this
to
be
a
community
project.
A
Let
you
know
you
tell
us
what
we
can
do
and
we'll
help
you
make
it
happen,
and
now
here
you
are
some
four
years
later.
This
is
what
the
world
looks
like
with
every
one
of
these
colored
slices
being
a
different
company
and
for
those
that
are
curious,
this
is
Google,
and
this
is
not
goo
and
that's
really
really
positive
sign.
When
you
can
say
when
that
line
goes
over
the
middle,
there
you're
doing
good
things.
A
That
means
more
people,
not
at
Google
work
on
kubernetes
and
contribute
upstream
than
the
people
who
are
working,
that
the
people
who
work
at
Google.
So,
let's
see
how's
cute
flow
well,
this
is
what
our
chart
looks
like
and
for
to
make
it
simple.
This
is
Google,
and
this
is
not
Google.
That's
after
one
year,
we
couldn't
be
more
ecstatic
about
this
and
I
want
to
make
that
orange
pie
as
big
as
possible.
A
We
want
to
hear
from
you
we
want
to
hear
from
the
community
about
what
we
can
do
to
make
cute
Flo
community
driven
to
solve
real
problems
across
the
industry,
to
specify
and
I
want
to
call
out-
and
this
is
just
a
small
set
of
people,
because
I
don't
have
that
long.
We've
gotten
really
significant
contributions
from
non
non-google
to
the
overall
platform
at
a
tube.
A
It's
a
micro,
service-oriented
architecture
that
Jay
will
be
demonstrating
just
shortly
for
doing
hyper,
parameter,
optimization,
really
really
cool,
and
in
kubernetes
one
or
a
cube,
Flo
0.3
we
introduced
a
brand-new
higher-level
construct,
called
a
study
job,
meaning
you
can
now
do
hyper
parameter
tuning
without
any
coding
whatsoever
and
we'll
demonstrate
that
in
just
a
minute,
second
I
want
to
call
it
tensor
RT.
This
is
a
serving
framework.
A
If
you
were
in
the
room
two
sessions
ago,
they
they
gave
a
great
deep
analysis
of
it
and
you
could
do
all
kinds
of
really
advanced
thing:
multiple
models
per
GPU.
You
can
support
heterogeneous
GPUs
and
things
like
that,
and
it
integrates
into
those
orchestration
systems
that
use
kubernetes
under
the
hood.
A
third
one:
we've
had
a
workflow
engine,
excuse-me
built
into
cube
flow
for
overnight,
or
excuse
me
over
12
months.
At
this
point
we
introduced
it
at
the
beginning
of
the
year.
This
is
from
Argo
and
you
have
workflow.
A
You
have
engines
and
pipelines,
and
we'll
show
more
about
this
in
a
minute.
We
also
have
Argos
CD
for
using
ml,
get
ops
where
you're
able
to
make
changes
in
code,
push
that
into
your
github
repository
and
have
that
cascade
through
and
make
changes
all
the
way
down
your
ml
pipeline
and
again,
you
know
I
know
I,
don't
have
an
enormous
amount
of
time
I.
A
So
now
in
a
Jupiter
notebook,
you're
able
to
spin
up
all
the
necessary
dependencies,
persistent
volumes
and
so
on
from
that
kubernetes
underlying
engine
Selden
has
a
very
rich
serving
framework
for
non
tensorflow
models
and
they've
built
both
a
commercial
and
an
open-source
offering
and
cube
bench
from
Cisco
who
does
ml
benchmarking
on
top
of
kubernetes
and
cue
flow,
and
makes
it
very
easy
to
explore
if
your
deployment
is
operating
at
its
top
performance.
So
with
that
introducing
their
cue
flow
0.4,
almost
it's
coming
out
next
week
and
the
reality
is,
is
here.
A
You
know
we've
heard
from
the
community
that
one
of
the
most
challenging
things
with
cue
flow
today
is
that
it
is
still
too
hard
to
use.
We
surface
a
lot
of
things
around
templating
and
deployments,
and
things
like
that
that
really,
as
a
data
scientist
or
an
m/l
engineer,
you
shouldn't
care
about.
So
we've
made
some
some
big
improvements
around
deployment,
with
both
a
simplified
setup
in
an
application
CRD,
and
then
also
we've
made
big
improvements
with
with
one
of
the
biggest
announcements
in
in
the
entire
year.
A
B
Thank
you
very
much
David,
so
we
have
a
lot
of
fun
things
to
show
you
today
regarding
cube
flow.
One
is
click
to
deploy
common
thing.
We
hear.
Is
it's
too
hard
to
install
cube
flow
I,
don't
want
to
have
to
deal
with
the
command
line.
I
don't
want
to
have
to
figure
out
how
to
install
case
on
it,
deploy
a
case
on
adapt
solution,
a
one-click
installation
tool
available
via
a
clean
web
interface.
How
did
we
do
that?
B
Well,
we
created
the
click
to
deploy
which
uses
a
bootstrap
container
and
the
KF
CTL
shell,
which
is
something
that
you
can
run
on
your
own
infrastructure,
which
worth
pointing
out.
It
will
install
cube
flow
for
you
pretty
automatically
automagically,
as
some
of
us
would
like
to
say,
it
also
enables
use
of
declarative
infrastructure
deployment
or
on
GCP.
We
call
that
the
deployment
manager
and
there
is
no
templating
tool
needed,
but
you
know
what
let's
show,
and
rather
than
tell
ever
for
me
one
minute
all
right.
B
So
this
is
what
it
would
look
like,
and
one
thing
I
want
to
call
out
real
quickly.
Is
that,
while
this
is
a
web
interface
that
is
for
Google
cloud
platform,
the
underlying
tools,
such
as
the
KF
CTL,
can
be
run
on
any
environment?
It's
not
just
a
Google
thing,
though,
though
this
web
interface
is
today,
as
you
can
see,
we
only
take
six
inputs
project,
name,
the
name,
your
deployment,
there's
a
web,
app
client,
ID
and
a
client
secret.
You
can
get
it.
We
give
you
detailed
instructions
on
how
to
acquire
those.
B
B
Jump
over
all
right,
so
jump
back
a
little
bit
all
right.
So
here's
our
deployment,
let's
take
a
look,
wait,
there's
no
workloads
yet!
Ok,
it
still.
Maybe
it's
still
building
a
little
bit
still
building
still
building.
Let's
refresh
well
look
at
there,
we
have
everything
that
is
bundled
within
cube
flow
that
you
used
to
have
to
do
a
lot
of
templating
with
case
on
it
deploying
the
case
on
application.
A
lot
of
manual
work
all
done
with
six
parameters
in
a
web
interface.
B
B
Well,
once
you
have
deployed
it,
what
happens
when
you
want
to
upgrade
I
mean
we
just
talked
about
point
four,
which
has
come
out
next
week
and
you
say
well:
I
just
deployed
point
I
upgraded.
This
was
great
to
install,
but
well
we
were
looking
at
get
good
ops.
How
do
you
maintain
the
cluster?
Maintaining
a
cluster
can
be
hard,
we've
implemented,
get
off
switched
I
should
point
out
was
a
term
coined
by
we
works,
get
ops,
driven
solution
to
manage
infrastructure
and
cut
the
cluster
code.
B
We
are
using
Argos
CD
and
they
actually
have
a
pretty
cool
demo
downstairs
if
you
haven't
taken
a
taking
it's
the
time
to
stop
by
the
booth
synchronize
the
cube
flow
deployment
with
a
git
repository,
and
we
have
a
blog
post.
That
shows
you
exactly
how
to
do
that.
If
you
go
to,
or
rather
our
documentation,
if
you
go
there,
you
can
follow
the
instructions
and
do
that
just
as
well
now
one
thing
we
just
announced
and
that
David
talked
about
Wars
cube
flow
pipelines.
The
problem
and
most
solutions
are
often
multistage.
B
Anybody
here
do
anything
with
ml
models
yeah.
So
any
anybody
here
will
agree
with
me
that
they're
off
in
multi
part
they're,
not
just
one
thing.
So
the
solution
was
to
leverage
micro
services.
Leverage
containers
make
each
stage
its
own
container
its
who
enable
code,
reusable
components
and
workflow
orchestration,
so
cute
flow
pipelines.
It's
a
Python
SDK
that
allows
you
to
scribe
and
containerize
the
different
tasks.
It
runs
on
our
go
and
offers
experiment,
logging
and
analytics,
and
each
step
is
containerized
and
lets
you
extend
it
as
needed.
B
B
Again,
let's
take
a
good
look
at
how
it
works,
so
here
we're
in
a
shell
we
are
going
to
playing
yes,
so
we're
going
to
go
ahead
and
take
a
look
at
the
Python
code.
Now
this
is
standard
Python
Python.
Anybody
here
who
knows
Python
can
read
this
and
it
will
show
you
exactly
how
to
define
the
workflow
each
each
step.
Here
is
a
container:
let's
go
ahead
and
package
it
by
executing
the
pipeline,
the
Python
code.
There.
B
Examples
are
very
beneficial
in
learning,
so
we're
going
to
take
a
look
at
this
taps
taxi
tip
example.
Here
now
you
can
see
these
different
steps
here.
These
different
squares,
each
one
is
a
container.
Each
one
is
a
process.
If
we
go
ahead
and
click
on
any
one
of
them,
we
will
see
that
it
has
input
parameters,
output
parameters,
you'll
see
that
each
one
has
slightly
different
ones
because
they're
again,
each
stage
has
its
own
container
they're
their
own
separate
process,
and
why
don't
we
go
ahead
and
upload
the
amol
we
just
created.
B
And
oh
there
you
go
simple
as
that,
and
now
we
have
two
stages:
training
and
post-processing.
Each
states
its
own
is
its
own
container.
We
have
input
output
parameters
and,
just
like
that,
you
were
able
to
create
a
multi
stage
pipeline
very
minimal
code.
You
created
a
Python,
you
create
Python
code
generator,
the
yam
will
uploaded
it
you're
good
to
go
so,
let's
go
ahead
and
start
an
experiment,
we're
gonna
name
it
pipeline.
Keep
things
simple.
B
B
Auto
scaling
now
here
we
have
a
conversation
going
on
between
the
data
scientists
and
IT
ops.
Tell
me
now.
If
anybody
recognizes
this
as
we
go
forward,
model
works
great,
but
I
need
six
mode
nodes.
Sure
thing
can
I
get
to
it.
After
all,
the
hundred
things
I
have
to
do
today,
all
right.
Well,
when
you
have
time
time
passes
time,
passes
time
passes.
Oh
IT
ops
eventually
gets
to
it
great.
We
have
six
nodes.
B
B
This
is
a
very
common
thing,
so
we
introduced
auto-scaling,
so
you
can
describe
the
job
and
kubernetes
will
take
care
of
the
rest.
So
you
describe
how
much
CPU
you'll
need
how
much
RAM
you'll
need
accelerators,
GPUs,
TP
use,
TF
jobs
will
then
delete
themselves
when
they're
finished,
so
the
node
pool
will
automatically
scale
back
and
you
can
cap
it
to
a
to
a
maximum
parameter.
So
that
way
it
doesn't
scale
to
an
infinity
your
bill
skills
to
infinity.
B
B
Refresh
job
didn't
run
wonder
what's
wrong.
Let's
take
a
look
alright,
so
we
have
three
nodes.
Let's
take
a
look
at
the
workloads.
Oh
wait!
My
GPU,
it's
unschedulable
what's
happening.
Three
nodes
are
available:
Oh,
insufficient
GPU
power.
That's
not
good!
Let's
go
back
to
the
cluster,
see!
What's
going
on.
Let's
take
a
look
at
some
of
the
nodes.
Okay,
so
wait.
We
have
four
nodes.
Now
one
just
came
up:
it's
not
available,
yet
all
right,
let's
refresh
see,
what's
happening
here
and
as
you
can
see,
the
nodes
is
being
prepared.
B
It's
now
in
this
not
ready
state,
but
it
went
from
red
to
yellow
and
now
that
we
refresh
again,
oh
that
fourth
node
is
ready
with
the
GPUs
that
we
need
to
run.
So
when
we
go
and
refresh
the
code
there
we
go,
it
ran
simple
as
that
it's
scaled
up
to
what
it
needed
to
do
are
scaled
up
to
the
resources
that
it
needs
so
that
it
could
do
what
it
needed
to
do,
which
is
a
good
thing
and
it
will
scale
down
when
it's
done
pretty
interesting.
B
Isn't
it
being
able
to
scale
up
to
the
resource
you
need
for
training,
then
scale
backed
out
I
mean
this
is
the
beauty
of
cloud
native.
This
is
the
beauty
of
the
cloud.
This
is
a
beautiful
cube
flow
being
able
to
use
resources
when
they
are
needed,
but
they're
not
having
to
have
the
perpetually
and
eating
that
cost.
B
So
we
talked
a
little
bit
about
cat
tip,
which
does
which
does
the
hyper
parameter
tuning.
It
uses
a
feature
called
study
job,
so
let's
go
ahead
and
take
a
quick
look
at
that
and
let's
see
here,
oh
here
we
go
so
we're
gonna
go
ahead
and
apply
this
example
file
to
build
out
the
nodes
that
we
need
and
the
job.
So
now
we're
gonna
get
the
study
job.
Ok,
there
we
go.
Let's
take
a
look
and
describe
it.
B
Ok,
so
this
looks
pretty
standard
for
anybody.
Who's
used.
Kubernetes
has
given
us
all
the
information,
so,
let's
scroll
up
a
little
bit.
Ok,
so
we're
using
see,
we
were
gonna
have
a
learning
rate
of
about
see.
We
got
points
your
warm
points
home.
Then
we
have
scrolling
down
a
little
bit.
Okay,
here's
our
pod.
B
Alright,
so
the
pods
being
built,
and
now
we
will
go
ahead
and
see
oh,
wait,
we're
seeing
the
exact
same
thing
we
saw
before,
though,
but
it's
it's
in
process
of
being
created,
get
the
notes.
Look
at
there.
We
have
three
nodes
ready
to
go
for
the
hyper
parameter.
Tuning
hyper
parameter
tuning
simply
will
spin
up
multiple
experiments
simultaneously
to
do
your
testing.
B
B
That's
exactly
what
it
is,
a
graphic
that
will
give
you
the
visualization
of
all
those
simultaneous
models
that
are
running
all
that
saw
simultaneous
training
that
is
going
on,
and
why
don't
we
skip
ahead
yeah
there
we
go,
so
you
also
have
this
option
here.
Where
you
get
another
graph,
you
can
also
create
some
visualizations.
We
all
like
sharing
graphs,
don't
we
you
can
compare
some
some
information.
B
B
Let's
group,
by
a
number
of
layers,
so
what
you're
seeing
here
is
you
have
this?
You
have
this
graph
power
by
cat
tube
and
I,
keep
wanting
to
call
a
ketubah
cat
tube.
Let's
go
ahead
and
and
model
DB,
let's
go
ahead
and
create
another
pipeline,
we'll
just
clone
the
exact
one
that
we
had
before
this
time.
We're
gonna
change
the
learning
rate,
we're
gonna,
add
a
layer,
we're
gonna
change
the
optimizer,
we're
gonna,
create
it.
It's
gonna
go
through
that
same
process.
It's
going
to
run
faster.
B
B
B
B
Well,
so
now
we
showed
you
some
of
these
pretty
interesting
features.
Let's
talk
a
little
bit
about
the
road
map.
Let's
talk
about
the
future
of
cut
cube
flow.
We
talked
about
where
we've
been.
Let's
talk
about
where
we're
going
so
we're,
we
are
continually
improving
Enterprise
readiness,
while
we
do
see
some
people
using
this
in
a
semi
production
type
environment.
You
know
we
always
want
to
improve
it
so
that
it
will
be
more
I,
get
ready
for
the
enterprise,
so
cleaner
upgrades,
our
back
compatibility
so
on
and
so
forth.
B
We're
getting
ridging
closer
to
version,
100,
Jupiter,
notebook
integration.
We
want
to
include
pipeline
experiment
comparison
and
model
management,
but
of
course
one
thing
that
we
always
believe
in
the
world
of
open
source
is
sometimes
the
best
feature
is
the
one
that
we
haven't
heard
of
and
the
one
that
you
tell
us
about.
So
you
can
tell
us
what
you
want,
or
you
can
help
we
can
always
use
more
coders.
We
can
always
use
more
more
ingenious
ideas
if
you
will
and
it's
a
whole
new
world.
B
You
know
we
talked
earlier
about
how
machine
learning
is
becoming
a
thing.
Data
science
is
increasing,
and
data
science
is
starting
to
touch
every
industry,
whether
it
is
I'm
searching
on
Google
for
a
pizza
place
in
a
new
city
or
Netflix
is
telling
me
what
videos
I
might
like,
based
on
the
videos
I
buy
before
Amazon's
telling
me
what
perk
products
I
purchased
before
data
science
and
start
to
expand,
and
we
can't
ask
people
to
become
a
PhD
in
statistics.
B
However,
we
can,
we
do
have
to
ask
ourselves:
how
can
we
help
people
enterance
with
a
spray
of
new
world
and
if
we
can
simplify
the
process
of
getting
started
and
of
building
out
this
layer,
take
care
of
all
of
that
busy
busy
work?
If
you
will
of
scaling
out
if
resources
acquiring
servers,
putting
servers
interacts
whatever
the
case
might
be,
we
are
able
to
better
or
better
provide
this
environment
for
people
to
start
entering
the
data
science
field
and
help
people
who
are
already
in
the
data
science
field
cube
flow
is
open.
B
We
have
an
open
community,
we
open
design,
open
source
open
to
ideas,
so
maybe
you're
asking
yourself
right
now.
Well,
how
do
I
partake?
How
do
I
actually
help
cube
flow?
Well,
cube
flow
org
forcus
on
github
check
us
out
on
slack,
get
us
up
on
twitter
talk
to
us
while
we're
here
we
have
quite
a
few
people
in
the
room
who
are
representing
the
cube
flow
community.
We
have
people
downstairs
in
the
Expo
Center,
so
there
are
many
ways
for
you
to
get
involved
and
I.
C
A
Yeah
absolutely
so
the
question
was:
are
we
thinking
about
Jupiter
lab
extensions
to
think
about
controlling
cube
flow?
Absolutely,
that's!
Probably
one
of
the
top
requests
that
we
have
data
scientists
obviously
know
Jupiter
really
really
well,
and
we
want
to
deeply
integrate
that
under
the
hoods
honestly,
honestly,
it
required
pipelines
landing
which
just
landed
at
the
beginning
of
November.
In
order
to
do
that,
because
oftentimes,
a
very
standard
path
would
be
okay,
I've
completed
my
notebook
click,
a
button,
take
that
notebook
package
it
into
a
distributed
container
and
then
run
it
as
a
distributed
job.
A
A
The
question
was
how
about
a
model
hosting
service
as
well
as
integration
into
K
native
and
other
solutions?
Can
you
say
more
about
what
you
mean
by
a
model
hosting
like
experiment,
tracking
overtime
or
external
yeah
yeah?
So
that's
one
of
the
reasons
we
built
on
kubernetes
and
and
had
this
microservices
first
architecture.
A
Not
everything
will
run
in
Q
flow
I
guarantee
you
whether
or
not
it's
your
data,
whether
or
not
it's
your
storage,
whether
or
not
it's
your
serving
that
everybody
who
is
doing
ml
today
is
piecing
together
services
from
across
the
board
and
because
we
built
on
kubernetes
and
because
it
is
a
micro
services,
first
architecture
it
doesn't
matter
a
pipeline
step
can
call
something
inside
of
Q
flow
running
on
the
same
crew,
Burnette
ease
of
deployment,
it
can
be
calling
a
second
kubernetes
deployment.
It
be,
can
be
calling
a
software-as-a-service
api.
A
It
doesn't
matter
as
long
as
that
service
knows
how
to
receive
a
you
know.
Web
hook
and
then
hand
back
to
pipelines
an
event.
We're
good
and
pipelines
is
very
rich
when
it
comes
to
that
simultaneously.
You
asked
about
K.
Native
again
same
deal
applies
the
run.
You
know,
think
of
K
native
as
just
the
underlying
runtime
for
executing
an
arbitrary
stage.
A
There's
absolutely
nothing
you
saw
here
that
is,
requires
a
full
node
or
requires
anything
of
that
kind.
You
could
very
trivially
say:
oh
okay,
I'm
going
to
hand
it
to
this
I'm
gonna
hand
this
container
to
K
native
under
the
hood
and
go
to
town.
We
don't
have
that
integration
yet,
but
we're
absolutely
talking
to
that
team.
C
A
So
that's
actually
the
the
follow-up
question
was
where's.
The
model
metadata
stored
experiment
tracking
is
a
thing
that
we're
trying
to
sort
out
right
now
we
would
love
engagement
from
the
community.
You
saw
there
that,
for
example,
Kath
Tim
will
track
all
the
different
type
of
parameter,
searches
and
things
like
that.
Give
you
results.
We
are
actively
talking
about
that
and
it
is
something
we
would
very
much
like
community
feedback
about
what
the
right
solution
is
there.
A
We
we
haven't
seen
anything
yet
that
is
exactly
the
right
fit
for
the
way.
Kubernetes
works,
but
that's
not
to
say
we're,
not
open
and
we're
happy
to
kind
of
like
land
wherever
that
goes
to
be
to
be
clear
pipelines,
as
you
saw,
offers
the
ability
to
track
multiple
pipelines
over
time
and
we're
planning
to
make
that
much
much
richer.
But
we
want
to
make
sure
that
we're
balancing
everything
properly
so
please
come
in
and
give
us
your
feedback.
C
A
That's
amazing,
that
is
amazing,
so
this
gentleman
over
here
said
that
the
acumen
project,
which
is
the
Linux
Foundation
deep
learning
project,
will
be
based
integrating
coop
flow
and
having
a
solution
come
shortly.
Please
do
check
in
with
the
KF
CTL
folks
and
the
bootstrap
container
folks,
just
to
see
cuz.
There
may
be
tools
that
you
can
reuse
as
it
stands
right
now
it.
What
you
describe
is
awesome.
It
should
be
very,
very
thin
the
amount
of
work
that
you
have
to
do.
Okay,.
C
C
B
A
The
question
was:
how
does
this
work
with
multiple
users,
multi-tenancy
is,
is
something
that
we're
trying
to
sort
through
right
now?
To
be
perfectly
frank,
what
you
see
integrated
into
coop
flow
today
is
Jupiter
hub,
which
does
has
the
cons
does
have
the
concept
of
multiple
users
and
it
depends
on
your
level
of
multi-tenancy
requirements
in
order
to
get
things
going
given
how
early
this
is
and
given
the
low
resource
requirements
of
a
kind
of
you
know
resting
cube
flow
deployment.
A
Probably
what
I
would
recommend
today
is
multiple
clusters,
but
again
that's
largely
dependent
on
how
hard
it
is
for
you
to
set
up
a
question
things
like
that:
most
clouds.
Obviously
it's
one
click
away,
and
then
you
just
hit
a
deployment,
but
the
reality
is.
We
have
heard
many
many
times
like
look
I
just
want
to
set
up.
One
cube
flow
I
want
to.
A
You
know,
have
a
web-based
API
that
understands
the
identity
of
my
users
and
things
like
that
and
and
lets
them
each
share
and
and
build,
and
things
like
that,
we
don't
I
wish
I
could
say
we
had
a
rich
solution
for
that.
But
please
come
to
the
community,
tell
us
your
requirements
and
and
I
promise,
if
it's
not
already
on
the
backlog,
we'd
love
to
add
it.