►
Description
For more Continuous Delivery Foundation content, check out our blog: https://cd.foundation/blog/
MLOps with Spinnaker: through the eyes of a beginner - Shivay Lamba, Meilisearch
Today Machine learning can be seen in literally every aspect in life. Machine learning helps us make good decisions by learning from the existing data models and applying good predictions of the next output to give us suggestions and feedback over existing solutions.
Machine learning does differ from standard software development in terms of how the notebooks are processed. Thus the question arises about how to deploy production ready machine learning?
What all different strategies currently exist through which machine learning can also achieve CI/CD. How can we use Spinnaker to enable that.
To get answer of all these questions, please attend this lightning talk!
A
So
welcome
everyone,
I
know,
I'm
standing
in
between
the
lunch
and
the
last
talk,
but
I'll
be
very
quick
because
it's
a
five
minute
lightning
talk.
I
was
originally
supposed
to
be
presenting
the
stock
at
open
source,
Summit
Europe,
but
unfortunately
I
couldn't
make
it
so
I'm
presenting
it
here
in
Detroit,
so
really
excited
for
it.
A
So,
as
you
can
see,
the
title
is
envelopes
with
Spinnaker,
so
we'll
see
how
you
can
leverage
the
use
of
us
continuous
delivery
platform
like
Spinnaker
in
in
the
entire
mlops
Tech
stack
so
really
excited
for
presenting
this
I'm
shivaya
I'm
a
developerations
engineer
at
milliseearch.
It's
a
rust
based
search
engine
and
you
can
connect
with
me
on
my
Twitter
at
how
develop
later
on
and
in
case
you
have
any
questions.
I'd
love
to
catch
up
with
you
at
lunch,
but
just
to
get
get
started
right.
A
So
initially,
let's
probably
take
a
look
at
this
really
expensive,
like
a
diagram
and
don't
worry,
I'll
be
breaking
this
down
into
like
a
number
of
steps,
but
this
kind
of
showcases,
the
entire
CI
CD
pipelines
in
the
mlops
life
cycle
life
cycle.
So,
just
to
kind
of
reiterate
what
exactly
is
mlops,
it's
machine
learning,
plus
devops,
and
if
you
take
a
look
at
any
example
of
a
very
simple
example
of
a
machine
learning
life
cycle,
you
start
with
the
training
of
your
data
right.
A
So
first
you
get
your
data
and
you
pre-process
your
data.
Then
you
start
to
train
your
data
over
existing
machine
learning
model.
And,
finally,
once
you
are
done
with
the
training,
you
evaluate
your
model
and
you
publish
it
and
then,
of
course,
once
you
push
it
into
production,
it's
really
important
to
kind
of
monitor
the
health
of
your
machine
learning
model
and
that's
where
you'll
find
the
usage
of
a
conditional
integration
and
deployment
and
like
in
the
entire
machine
learning
MN
Ops
Tech
stack.
A
A
A
The
next
step
is
that
you
want
to
basically
create
a
machine
learning
pipeline,
so
an
ml
pipeline
can
be
defined
as
a
step
number
of
steps
that
you
can
take
to
basically
process
your
entire
end-to-end
machine
learning,
work,
workflows
right
right
from
getting
your
data
to
getting
your
production
results.
So
there
are
multiple
ways
in
which
you
can
run
your
ml
pipelines.
So
there
are
tools
like
Cube
flow
flight
and
even
like
you
know,
Selden
so
I
mean
there
are
a
number
of
ways
you
can
do
that.
A
But,
for
example,
we
will
take
a
cube
flow,
which
is
a
very
popular
open
source
tool
and
you
can
use
the
cube
flow
pipelines
to
have
your
entire
machine
learning.
Tech
stack
be
defined
with,
like
you
know,
each
and
every
individual
step
as
a
separate
component
within
this
pipeline.
Now
this
is
where
we'll
start
to
integrate
our
CI
CD.
A
A
So
if
you
are
having
multiple
deployments
and
you
want
to
roll
back
to
one
of
your
existing
deployments,
you
can
use
Spinnaker
and
it,
like
you,
know
the
Spinnaker
UI
lets
you
do
that
very
easily,
and
so
you
can
manage
your
deployments
and
you
can
also
schedule
your
deployments
as
well,
based
on
certain
events
or
based
on,
let's
say
a
Cron
job
that
you
might
want
to
run
every
week
right.
So
an
example
of
that
could
be
that
you're
collecting
your
data
at
the
end
of
every
week.
A
It
could
be
like,
let's
say,
a
historical
data
and
you're
retraining
your
model
from
time
to
time,
so
you
can
do
all
of
that
and
manage
that
with
the
help
of
Spinnaker.
And
of
course
this
is
what
it
would
look
like
the
pipeline.
So
at
the
right
at
the
leftmost
corner,
you
have
the
configuration.
So
this
is
where
you're
defining
the
yaml
file,
which
describes
the
entire
machine
learning
pipeline.
So
over
here
you
will
Define.
A
What's
your
POD
containers
and
the
type
of
Machinery
model
that
you
are
trying
to
train,
and
then
you
you'll
run
some
tests
with
the
help
of
Jenkins
and
then
you'll
actually
run
your
machine
learning
Pipeline
and
then,
once
if
the
machine
learning
pipeline
is
successfully
run,
you
can
go
ahead
and
deploy
that
with
the
help
of
Spinnaker.
And
if
you
take
a
look
at
in
terms
of
like
the
model
serving
so
what
exactly
is
model
serving?
A
Let's
say
whenever
there
are
some
changes
that
will
need
to
take
place
inside
of
your
model,
and
this
is
really
prevalent
when
you're,
like
you
know,
going
ahead
and
making
some
changes
to
your
model
and,
let's
say
it's
into
production
and
based
on
some
of
the
production
data.
There
might
be
situations
where,
like
you
know,
there's
some
unexpected
data
that
comes
in
and
your
model
is
not
performing
very
well
so
from
time
to
time,
as
your
model
goes
into
production.
A
You
have
to
make
some
incremental
changes
in
your
model,
so
that's
also
something
that
you
can
do
with
the
help
of
Spinnaker.
It's
a
model,
delivery
and
model
serving
becomes
very
easy,
with
the
help
of
Spinnaker
as
well
and
in
the
future.
If
you
want
to
also
monitor
the
health
of
your
model,
so
get
some
Telemetry
data
and
how
the
model
is
performing,
you
can
use
tools
like
Prometheus
grafana
dashboards,
to
kind
of
visualize.