►
From YouTube: Eco-Friendly ML: How the Kubeflow Ecosystem Bootstrapped Itself - Peter MacKinnon, Red Hat
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
Eco-Friendly ML: How the Kubeflow Ecosystem Bootstrapped Itself - Peter MacKinnon, Red Hat
How do you bootstrap an open source project that aims to provide stability, composability, and portability for machine learning? You use Kubernetes and its rich ecosystem to implement the pieces of infrastructure that you need to deliver a comprehensive ML platform for data scientists and DevOps engineers alike. This talk will explore the various integrations that have enabled Kubeflow to quickly emerge as the de-facto machine learning toolkit for Kubernetes. We'll look in detail at not only how Kubeflow leverages Ambassador, Argo, Ksonnet, and JupyterHub, but also examine integration with complementary projects such as Pachyderm and SeldonIO. You will leave this talk with a better understanding and inspiration of how a particular project can rapidly achieve its potential by working with other projects, and that those inter-project collaborations enrich the entire Kubernetes community.
To learn more: https://sched.co/GrTc
A
Welcome
to
this
talk,
which
is
called
eco-friendly
ml,
by
the
way
my
dad
and
I
used
to
sail,
we
would
probably
never
take
the
boat
out
and
that
kind
of
swell
there.
So
I
don't
know
if
anyone
else
is
noticed
that
this
is
a
community
talk
like
a
lot
of
the
other
talks
in
this
in
this
venue.
So
if
you're
coming
for
you
know
deep
detail
on
coop
flow,
we
have
lots
of
other
great
talks.
Michele
here
is
an
a
Mia.
A
It
had
a
great
one
yesterday
and
and
we're
gonna
have
birds
of
a
feather
later
on,
but
this
is
a
community
talk,
so
it's
going
to
be
very
light
on
details
and
we're
gonna
basically
talk
about
in
general,
the
machine
learning,
DevOps
problem
space
and
with
respect
to
coop
flow,
what
the
project
goals
and
aspirations
the
original
ones
and
what
they
continue
to
be,
and
then
how
the
community
and
project
started
to
come
together.
Basically
how
the
pieces
started
to
fall
into
place
and
towards
the
end,
we'll
talk
about
in
general,
how
good
open-source
works?
A
So
the
starting
point
is
that
today
it's
pretty
difficult
to
do
big
machine
learning
when
you
think
of
all
the
different
types
of
components
and
pieces
that
go
into
a
machine
learning,
workflow
and
pipeline.
You
have
your
data
ingestion
data
analysis.
You
go
do
feature
extraction.
You
have
data
scientists
working
with
notebooks
and
they're,
putting
together
models.
Those
models
need
to
be
deployed
and
served,
and
in
some
cases,
maybe
you're
reliant
on
GPUs
for
parallel
processing.
There's
lots
of
different
factors
that
come
into
play
when
you're
dealing
with
the
pipeline
for
machine
learning.
A
So
so
the
concept
here
with
the
coop
flow
project
in
the
community
was
to
bring
machine
learning
for
for
everyone.
Basically,
but
as
I
mentioned,
there's
a
lot
of
pieces
are
required
to
be
in
place
to
have
a
comprehensive
machine
learning
platform
and
it's
really
difficult
for
one
project
or
one's
an
institution
to
try
and
solve
that.
A
So
with
this
project,
the
idea
was
to
basically
start
with
sort
of
a
high-level
mission
statement,
and
some
of
the
goals
of
that
mission
are
to
make
it
portable.
So
whatever
we're
building
for
this
machine
learning
platform
should
be
able
to
go
from
bare
metal
out
to
the
cloud
scalable
from
one
developer,
one
machine
to
hundreds
and
it
should
be
composable
and
that's
the
modern
micro
service
architecture.
You
should
be
able
to
take
different
pieces
and
have
good
separation
of
concerns
and
have
those
well
integrated
together.
A
So
guess
what
kubernetes
the
conference
we're
at
it
checks
all
these
boxes.
It's
an
ideal
platform
for
doing
this
sort
of
deployment
or
composition
of
a
platform
for
machine
learning.
So
that's
sort
of
the
obvious
starting
point.
But
beyond
that,
where
do
you
go
so
so?
I
know
Jeremy's
here,
but
I
made
up
this
attribute
for
Jeremy
and
I
doubt
that
he
actually
ever
said
this.
But
the
reason
I
have
it.
A
This
project
that
you're
putting
together
so
just
like
Jerry
Maguire,
we
start
with
you,
know,
sort
of
a
more
specific
mission
statement
and
the
coop
flow
mission
is
dedicated
to
making
deployments
of
machine
learning
workflows
on
kubernetes,
simple,
portable
and
scalable.
Ok.
Now
that
second
point
is
really
important:
we're
not
trying
to
reinvent
the
wheel,
there's
lots
of
good
machine
learning
frameworks
that
already
exists,
so
tensorflow
scikit-learn
by
torch.
A
The
goal
of
this
project
is
not
to
reinvent
some
of
those
those
types
of
projects
and
try
and
improve
on
what
they're
doing
it's
more
to
integrate
the
best
of
breed
of
these
open-source
projects
and
incorporate
them
into
a
comprehensive
machine
learning
framework
so
and
then,
finally,
anywhere
you
can
run
kubernetes.
Ideally,
you
should
be
able
to
run
coop
flow.
Ok
so
that
that
was
sort
of
the
overall
mission
statement.
Great,
but
there's
still
lots
of
questions
and
issues.
So
how
do
we
run
our
data
data
science
notebooks?
A
What
about
the
data
pipelines
that
we
need
and
the
data
provenance
in
terms
of
you
know
what
what
actually
produced
this
model?
What
were
the
inputs
and
what?
What
outputs
do
we
have
in
terms
of
the
notebooks
themselves?
What
about
access
in
Grasse
control?
And
after
building
this
platform,
we
need
continuous
integration
and
can
continuous
deployment
so
there's
inference
deployment
in
two
different
environments.
There's
still
lots
of
questions
so,
with
the
coop
flow
project,
it
became
a
matter
of
sort
of
going
around
and
looking
at
what
was
out
there.
A
Ok,
what
are
the
components
that
we
could
start
piecing
together
to
sort
of
bring
this
project
into
reality
so
where
to
start-
and
you
know-
Craig
previously
talked
a
lot
about
soar-
this
this
notion
with
open
source
open,
is
key
in
terms
of
discussion.
So
when
you
go
to
blogs
or
forums,
conferences
like
this
there's
a
lot
of
good
discussion
and
that's
an
obvious
starting
point.
A
So,
when
you're
trying
to
build
something
like
Kubb
flow,
the
idea
is
to
maybe
reach
out
to
other
communities,
see
what
they're
working
on
do
their
goals
align
with
your
goals.
Do
they
have
a
feature
set
that
can
actually
help
you
solve
a
problem?
So
that's
a
good
starting
point.
What
happens
after
that?
It's
best
I
actually
start
participating
in
these
communities.
Okay,
so
the
best
these
communities
have
slack
channels
or
they're
on
Gitter.
A
We
have
a
fair
amount
of
interests
and
some
enthusiastic
collaborators,
but
one
point
I
wanted
to
make.
Is
it's
really
important
when
you're
interacting
with
these
projects,
there's
typically
a
main
line
of
development?
You
know
the
master
and
it's
important
that
that's
kept
pure
right
and
that's
how
most
communities
sort
of
run
their
their
code,
development
and
CI
and
CD.
So
sometimes
you
have
an
enthusiastic
contributor
who
gets
sort
of
tripped
up
by
the
notion
of
rebasing
and
squashing
commits,
and
things
like
that.
A
So
this
happened
a
couple
of
times
in
the
coop
flow
project
early
on
and
we
as
a
community
exercise
a
great
deal
of
patience
and
trying
to
to
help
that
individual
along,
because
what
they're
offering
was
valuable,
their
interest
and
enthusiasm
and
then
finally,
you
know
what
they
were
trying
to
implement.
But
you
know
you've
you've
got
to
work
around
some
of
the
issues
if
there's
sort
of
a
lack
of
understanding
of
some
of
these
concepts
with
with
get
in
other
SCMS.
A
So
an
obvious
one
that
we
probably
all
know
about
is
github
makes
it
really
easy
to
cross-reference
issues
across
repos
across
projects.
So
you
can
easily
sort
of
create
start
creating
sort
of
this
symbiotic
sort
of
connection
between
your
products
by
referencing
issues
in
both
places
and
then
finally
there
in
that
section
you
want
to
treat
that
community
with
the
same
respect
and
importance
that
you
feel
is
deserved
in
your
project
and
so
I
think
it's
a
good
sort
of
sort
of
stance.
They
have
basically,
when
you're
approaching
an
opportunity
to
integrate
with
another
project.
A
So
some
of
that
is
prescriptive.
What
I've
just
described
now
I'll
fall
back
into
sort
of
the
retrospective
part
of
this,
and
so
with
the
coop
flow
project.
We
talked
about
this
this
platform
that
we
needed
for
machine
learning
and
we
want
to
have
integration
with
all
these
other
different
projects,
but
we
needed
on-ramp.
A
We
need
a
point
of
integration
to
bring
these
other
projects
into
the
fold
and
the
case
on
a
project
turned
out
to
be
represents
art
of
that
on-ramp,
so
coop
flow
uses
caisson
it
it's
a
JSON
based
tool,
basically
to
generate
these
Cooper
net
deployment
objects,
so
you
express
them
basely
in
JSON
there
they
have
primitives,
basically
for
things
like
pods
and
services
and
and
CR
DS,
and
with
that
you
can
write
sort
of
very
descriptive
and
powerful
mechanisms
for
integration
with
the
components.
Okay.
A
Now
what
you
get
once
you
use
case
on
it
to
deploy
these
resources
is
what
you
would
generally
recognized
from
kubernetes
in
terms
of
Yambol
or
JSON
for
the
resource
objects.
The
other
powerful
part
of
case
on
it
is
that,
because
you
can
have
parameters
for
the
components,
you
can
basically
have
different
settings
for
different
environments.
Perhaps
you
have
a
dev
environment
or
test
environment
staging
prod.
You
can
have
different
tweaks
for
your
components
that
are
used
to
push
the
deployment
of
Kubb
flow
into
that
environment,
okay
and
also
continuous
deployment.
A
So
when
you
make
your
changes,
you
setup
your
parameters
and
do
an
apply.
You
are
basically
getting
a
regeneration
of
your
coupe
flow
deployment
and
by
that
same
token,
you
can
actually
tear
it
down
with
case
on
that
as
well.
So
you
can
have
a
deletion
of
the
resources
there
ingress.
That
was
one
of
the
other
questions
and
one
of
the
projects
we
attached
to
or
integrated
with
early
on
was
ambassador.
So
that's
a
cloud
native
ingress
controller
and
it's
based
on
the
very
popular
envoy
project.
A
This
is
a
theme
with
all
these
integrations
that
we
have
it's
cloud
native,
kubernetes
native
workflow
manager,
so
CRD
and
an
operator.
So
that's
an
important
piece
in
the
coop
flow
project.
It
it's
the
underpinning
for
our
pre-imposed,
submit
testing
it
actually
exercises
and
then
test
workflows
for
each
commit
that
is
going
into
Kubb
flow.
It
also
is
involved
or
part.
The
KU
bench
sub
project
that
was
donated
by
Cisco
and
also
the
KU
flow
pipeline
sub
project
makes
use
of
Fargo
as
well.
A
So,
basically,
you
define
your
workflows
in
y
mo
and
then
you
activate
them
and
what
you're
building
up
is
basically
a
dag,
a
directed
graph
of
these
workflows
to
basically
do
things
in
the
case
of
our
pre
and
post,
submit
testing,
standing
up
environments
and
basically
exercising
tests.
So
in
general
it's
also
really
good
for
just
doing
machine
learning
pipelines.
You
can
stitch
together
things
for
for
jobs,
for
doing
machine
learning
or
data
processing,
and
basically
putting
together
pipelines.
A
Selden
was
one
of
the
earliest
components
that
got
integrated
into
Kubb
flow
and
that's
a
project
that
deals
with
running
in
inference,
graphs,
deploying
and
manage
them
at
scale
on
kubernetes.
So
these
graphs
are
basically
deployed
as
micro
services
and
they
have
different
types
of
graphs
that
can
be
stitched
together.
Basically,
so
you
have
different
components
that
each
run
as
a
micro
service
model,
transform
or
router,
and
they
support
a
lot
of
the
different
popular
machine
learning.
A
Libraries
like
tensorflow
and
scikit-learn,
and
you
can
basically
mix
and
match
those
components
in
various
ways
as
g
RPC
and
rest
inch
interfaces
and
basically
it
uses
openshift
sui
to
generate
these
model.
Wrappers
that
get
deployed
into
the
the
graphs
that
are
run
for
inference
in
one
of
the
other
early
projects
that
got
involved
or
we
integrated
with
coop
flow
is
pachyderm
and
I
was
just
talking
to
Nick
Harvey
right
here
up
front
and
I.
A
Guess
you
guys
started
described
as
get
for
data
science,
I
think
that
still
applies
there.
So
the
idea
is
to
have
reap
experiments
and
you
construct
data
science
pipelines
and
there's
a
specification
in
JSON
for
expressing
the
transformation
and
the
degree
of
parallelism
and
the
types
of
inputs,
and
you
guys
for
the
open-source
version
of
your
product.
A
Now
these
days,
there's
companies
like
Intel
and
at
Red
Hat.
We
are
generating
some
wheel
files,
basically
that
have
different
optimizations
that
you
can
turn
on
for
tensorflow.
With
respect
to
the
the
coop
flow
project.
We
use
the
default
ones
that
are
produced
by
the
tensor
flow
project.
We've
also
worked
with
the
team
at
kaggle.
They
have
a
Python
notebook
and
we've
basically
adopted
and
adapted
that
to
run
within
coop
flow
natively.
So
that's
kind
of
different
from
our
curated
notebooks.
A
They
have
a
lot
more
content
in
the
gaggle
notebook.
It
makes
for
a
much
larger
image
size,
but
it's
still
a
very
powerful
tool
for
basically
having
a
whole
plethora
of
different
machine
learning.
Libraries
and
visualization
libraries
at
your
disposal
and
we've
also
been
working
with
the
rapids
AI
project
from
nvidia.
They
also
have
a
notebook
image
and
we've
been
adapting
that
into
coop
flow
as
well.
A
So
there's
been
some
give-and-take,
some
back
and
forth
with
that
that
team
about
the
adaptation
and
some
things
that
we
noticed
that
result
in
basically
issues
against
their
project,
but
it's
going
really
well
and
so
that's
sort
of
the
technical
side
of
things
everything
I
mentioned
up
to
there
in
terms
of
the
coop
flow
project
itself,
sort
of
on
the
the
non
technical
side.
We
have
very
good
project
governance,
so
we
have
a
code
of
conduct
and
that
caught
a
code
of
contact
is
very
much
in
line
with
CNC
F.
A
A
So
part
of
that
work
is
to
set
standards
in
place
for
a
healthy
and
positive
community,
so
the
use
of
welcoming
inclusive
language
respectful
of
differing
viewpoints.
You
know
the
give-and-take
of
constructive
criticism
and
basically,
if
everything
should
be
oriented
around
focusing
on
what's
best
for
the
community,
so
also
as
part
of
that
we
have
conflict
resolution
protocols
if
there's
a
contributor
or
some
other
project.
Member
member
who
wants
to
raise
an
issue
with
regards
to
project
governance,
we've
laid
out
very
clearly
the
steps
that
they
should
take
to
raise
awareness
about
their
concern.
A
So
this
is
what
we
do
in
the
coop
flow
project
and
we
hope
to
see
it.
You
know
from
the
other
projects
and
communities
that
we
interact
with
so
today,
the
community.
It's
on
github
its
ASL
223
repositories,
135
contributors
in
the
core
project
in
the
core
repo
and
as
I
wrote
this.
There
was
about
nine
hundred
nine
commits
in
that
core
repo,
so
we're
just
releasing
0.34
and
we
are
getting
ready
to
release
0.4
dot
zero,
so
we're
having
these
integrations
that
are
continuously
happening.
A
So
it's
you
know
this
sort
of
as
a
community
we're
trying
to
keep
up
with
a
tremendous
amount
of
interest
and
enthusiasm
that
people
have
shown
for
the
project
and
we've
had
a
lot
of
institutional
participation,
I'm,
not
meaning
to
leave
anyone
out
there
I
apologize
right
now.
If
anyone
feels
slighted.
These
are
some
of
the
the
companies
and
organizations
that
have
been
actively
involved
and
and
provided
valuable
contributions,
which
brings
us
to
the
power
of
open
source.
A
A
It's
often
been
said
that
it
was
just
a
year
ago
that
Kubb
flow
was
announced,
so
here
we
are
a
year
later
and
we've
made
great
strides,
but
there's
still
a
lot
more
work
to
be
done,
love
to
see
some
more
participation
in
contributions
from
people,
I'd
love
to
hear
more
about
your
use
cases.
But
the
idea
here
is
that
when
we
interact
with
your
community
and
your
community
interacts
with
ours,
everybody
wins
it's
a
win
all
around.
A
So
there's
free
exchange
of
concepts
in
code
and
it
works
to
really
promote
each
other's
projects
and
communities.
So
there's
a
real
multiplier
effect
that
is
in
effect
for
the
respective
communities
user
base.
We
can
all
sort
of
build
off
them
and
I
think
for
the
contributors.
Certainly
I
feel
this
way
is
that
there's
of
a
sense
of
empowerment?
You
know
many
of
us
have
had
day
jobs
where
we're
not
involved
in
open
source,
and
sometimes
it
can
be
a
slog
it's
different
with
open
source.
A
You
have
a
lot
of
people
who
are
interested
in
the
work,
you're
doing
and
they're
supportive
of
the
work
you're
doing
and
we
all
sort
of
celebrate
each
other's
victories
as
it
were
in
contributions
and
that's
it
I,
don't
know
how
I'm
doing
for
time,
but
that's
a
whirlwind
trip
through
the
coop
flow
community.
So
a
few
minutes
for
questions.
A
We
good
okay,
one
reminder
right
after
this
is
the
coop
flow
birds
of
a
feather
I,
don't
know
if
it's
right
after,
but
it's
3:40
I
think
it
starts
and
is
it
6:18
6:20
might
be
the
rooms,
yeah
and
we'd
love
to
see
you
all
there
we'd
love
to
hear
your
experiences
after
your
coop
flow
user
love
to
have
your
feedback
ideas
for
contributions,
definitely
use
cases
so
come
along
and
join
us
and
we'd
love
to
meet
you
there.
Okay,
thank
you.