►
From YouTube: CDF - SIG MLOps Meeting 2021-08-12 (2 of 2)
Description
For more Continuous Delivery Foundation content, check out our blog: https://cd.foundation/blog/
A
A
So,
have
you
been
along
to
one
of
these
sessions
before.
B
A
Okay,
great
well,
what
we're
doing
here
really
is
just
trying
to
coordinate
some
of
the
work
on
the
ml
ops
roadmap,
and
I
don't
know
if,
if
you've
had
a
look
at
the
roadmap
document,.
A
What
we're
doing
here
is
trying
to
collate
a
set
of
documentation
which
basically
paints
the
picture
of
all
of
the
challenges
that
you
can
expect
to
encounter
when
working
with
machine
learning
assets
in
production,
environments.
A
And
so
the
the
mlx
roadmap
document
sets
out
our
vision
for
what
ml
ops
should
be
within
a
breast
practice
environment.
A
And
it
proposes
a
set
of
technology
requirements
that
are
necessary
to
be
able
to
successfully
address
those
challenges,
and
then
it
it
also
tries
to
monitor
the
the
available
solutions
and
potential
solutions
out
there.
A
A
And
the
idea
really
is
to
encourage
teams
that
are
working
on
continuous
delivery
and
ml
op
solutions
to
actually
fully
understand
what
what
the
customer
problems
are
in
that
space
and
what
capabilities
we
need
in
in
the
tooling
overall,
so
that
people
can
actually
achieve
what
they
need
to
be
able
to
achieve.
B
Right
I've
been
interested
in
mlaps
for
some
time.
I
guess
this
year
I
have
some
background
machine
learning
and
that
kind
of
part,
but
now
I'm
moving
towards
the
cloud
and
much
more
developer
operations,
kind
of
side
of
things
and
that's
how
I
kind
of
found
out
about
the
cd
foundation
and
before
this
I
spent
some
time
looking
at
what
emulates
first,
I
think
there
was
a
conference
over
a
cdcon
and
there
was
a
presentation
over
as
well,
so
the
current
consumer
captured
my
interest
and
I've
been
thinking
like
as
a
developer.
B
I
would
like
to
be
so
sufficient
to
be
able
to
make
some
products
on
my
own
like
projects
right
and
in
corporate
machining
into
them.
So
that's
why
I
really
wanted
to
understand
how
machine
learning
operations
kind
of
work
together,
decelerating
all
technologies
out
there
like
what's
practices.
What
problems
do
you
guys
have?
B
So
that's
kind
of
my
like,
where
my
interest
comes
from
and
that's
kind
of
what
I'm
trying
to
look
ahead
and
learn
from
the
sick
here,
just
trying
to
like
understand
what
kind
of
problems
they
are
and
then,
and
if
there
is
some
way,
how
can
I
help
to
contribute
to
that
to
be
like?
If
I
can
share
my
knowledge,
you
if
I
can
help
with
my
time,
they're
pretty
good
as
well.
B
I'm
just
student,
I'm
I'm
in
the
last
year
of
my
university,
so
kind
of
exploring
this,
these
different
spaces
and
learning
about
these
and
these
kind
of
things,
because
open
source
is
really
something
I'm
interested
in.
So
that's
about
me.
A
Okay,
thank
you.
Well,
there's
there
there
are
lots
of
currently
unsolved
problems
in
this
space.
So
it's
a
it's
a
fascinating
area
if
you're
looking
for
something
to
get
your
teeth
into,
because
there
are
lots
and
lots
of
challenges
and
relatively
few
people
really
working
on
them.
A
B
So
I
I
have
a
question
about
envelopes
right.
How
is
this
different
from
data
engineering,
virtual
machine
learning
right?
I'm
expecting
that
you
have
a
component
for
machine
learning.
That's
somewhere
between
the
pipeline
that
you
have
with
data
engineering
like
you
have
some
ingress.
You
have
some
egress,
you
have
a
pipeline
and
then
you're
probably
doing
some
predictions
for
the
machining
component.
Is
that
right.
A
So
that's
that's
close
to
what
we
see
as
being
best
practice,
but
it's
actually
a
very
long
way
away
from
what
people
are
actually
doing
right
now.
A
So,
to
give
you
a
little
bit
of
background
a
lot
of
what
goes
on
in
the
machine
learning
space
at
the
moment
has
come
from
the
data
science
field
rather
than
from
the
sort
of
software
engineering
field,
so
in
in
that
space
you're.
Typically
looking
at
a
group
of
people
who
come
from
a
mathematical
background
and
studied.
B
Your
voice
is
cut
off.
I
cannot
hear
you.
B
B
B
I'm
not
sure
I
can
hear
myself
and
I
can
listen
to
audio
on
youtube.
I
check
the
zoom
and
it's
kind
of
recording
for
me.
I'm
not
really
sure
why
this
is
happening.
A
Okay
right,
I've
switched
to
an
alternate
microphone
must
be,
must
be
a
driver
problem.
A
So
yeah,
so
what
I
was
saying
was
that
mostly
or
the
majority
of
people
who
are
working
in
a
machine
learning
field
at
the
moment
have
come
from
an
academic
background
where
they've
studied
a
lot
of
machine
learning
techniques
and
statistics.
A
A
B
A
Now,
in
practice,
that's
not
a
very
good
strategy
for
productionizing
your
machine
learning
models,
because
it's
very
hard
to
make
it
consistent
and
reliable.
A
It's
hard
to
hard
to
make
it
actually
a
repeatable
process
with
versioning,
and
there
are
no
a
number
of
problems
that
you
can
run
into
if
you're
relying
on
you
know,
dupes
notebooks
as
the
way
of
managing
the
the
essential
parts
of
your
process.
B
A
And
then
there
are,
there
are
a
whole
bunch
of
broader
challenges
that
expand
out
from
from
that,
so
so
in
in
an
academic
situation.
Python
is
really
useful
because
it's
easy
to
learn
and
it's
got
lots
and
lots
of
machine
learning
libraries
associated
with
it.
B
A
Which
means
that
it's
inefficient
and
it's
also
very
difficult
to
manage
dependencies
in
a
python
environment.
A
B
That
makes
a
lot
of
sense
and
I
think
that's
pretty
important
as
well,
because
you
may
also
want
to
induce
testing
in
between
them
and
then
there's
also
the
concept
of
a
b
testing,
as
in
try
and
test
rfq
model
performs
the
same
as
you
did
some
months
ago,
and
it's
doing
it
right
now
and
maybe
it
may
be
making
some
predictions
that
you
may
personally
not
expect,
but
the
machine
learning
model
is
giving
out
and
you
may
want
to
test
like
the
data
operations
are
correct
or
not,
if
it
having
fed
into
that.
B
It's
not
anything.
Some
pilot
process
is
a
bit
like
trying
to
understand
what
data
was
used
to
create
a
kind
of
model
as
well
and
trying
to
make
sense
right.
Is
there
any
difference
in
data
that's
coming
in
the
future,
as
in
data
versioning
right
that
we
have
some
systems
over
there
like
we
have
where
source
code
versioning
would
get
so
your
bit
bucket
do
stuff
like
that.
So
I
think
that's
also
a
component
of
it,
because
I've
seen
people
treat
machine
learning
models
as
black
boxes,
where
you
just
given
some
data.
B
It
gives
you
some
output
and,
most
of
the
time
I've
seen
innovation
come
from
people
learn
to
do
feature
engineering
properly,
where
they
manage
to
extract
the
right
properties
and
the
features
from
a
given
data
set.
I
think
that's
where
the
most
value
I
think
comes
from
when
it
comes
to
deploying
models
and
testing
out
the
accuracy.
Obviously,
every
organization
would
want.
I
don't
want
to
do
the
best,
but
there's
100
machine
learning.
It's
never
really
possible.
A
Yeah
you're
absolutely
right.
There's
there
are
some
very
significant
challenges
and
in
fact
the
data
versioning
one
is
very
easy
to
say,
but
to
actually
implement.
It
is
incredibly
difficult
because,
if
you've,
if
you've
got
say,
10
petabytes
of
training
data,
then
the
practical
challenges
of
managing
versions.
Snapshots
of
that
much
data,
especially
if
it's
changing
significantly
on
a
on
a
daily
or
weekly
cadence.
A
B
Right,
so
so
what
so,
so?
What
does
eminent
always
propose
as
a
solution
right?
There
are
different
solutions,
some
like
coming
from,
because
I
also
have
an
interest
in
software
architecture
and
database
design
stuff
like
that
and
coming
from
that
perspective,
I
think
there's
always
a
trade-off
when
you
have
to
make
when
you're
working
with
such
a
data
set,
you
have
to
think
about
backups.
You
have
to
think
about
versioning,
you
think
about
if
everything
is
being
consistent
and
there's
also
property
of
assets,
if
you're
doing
such
kind
of
property.
B
If,
for
example,
you
have
an
application
that
has
some
kind
of
streaming
kind
of
machine
learning
where
you
say
like
facebook
right,
facebook
has
a
lot
of
different
mechanisms
built
into
it
and
some
functionalities
or
features
try
to
guess.
You
recommend
you
something
based
on
what
kind
of
drinks
like
things
are
you
liking
your
posting
or
stuff?
Like
that,
I
mean
I
should
recommend
you
the
same
kind
of
things
in
the
future
as
well.
B
So
there's
also
the
idea
where
you
have
to
continuously
get
data
from
the
user,
and
then
you
have
to
process
the
output
or
somewhere,
and
then
you
have
to
make
it
a
pipeline
in
such
a
way
right
where
you
had
a
machine
learning
model
and
learning
is
from
you
know,
new
data
that
is
all
coming
into
it
at
the
same
time
and
it
has
to
give
some
predictions
on
this
of
that
problem.
So
it's
a
very
changing
kind
of
model
that
you
have
to
think
about
as
well.
A
There
there
are
multiple
problems
in
in
this
space.
One
one
of
the
real
challenges
is
that.
A
You've
got
two
different
categories
of
machine
learning
if
you
like,
you've
got
the
type
of
machine
learning
that
is
run
offline
against
pre-classified
data,
where
it's
easy
to
put
that
through
some
sort
of
pipeline
and
manage
the
whole
process
end
to
end.
A
But
then
you've
also
got
a
class
of
learning
which
is
happening
online.
A
So
the
system
is
learning
from
its
inputs
in
real
time
and
is
retraining
itself
dynamically
on
the
fly
and
at
that
point,
you're
operating
outside
of
a
build
environment,
you're
in
your
operational
environment
and
and
you're
you're
training
in
production.
A
So
so
so
there's
this
this
there
are
major
differences
between
those
two
solutions
and
it's
it's
difficult
to
do
things
in
the
first
scenario
and
very
difficult
to
do
things
in
the
second
scenario,
and
so
in
a
lot
of
cases
right
now,
people
are
just
not
bothering
to
do
the
a
lot
of
the
basic
best
practice
to
to
you
know
properly
test
and
audit
things.
A
So
it's
it's
the
wild
west
out
there
in
in
many
cases
right
now,
there's
the
quality
controls
are,
are
limited
and
we're
likely
to
see
significant
numbers
of
large
incidents
over
the
next
few
years.
As
you
know,
some
of
these
problems
start
to
turn
into
issues.
B
So
I
can
see
people
from
software
engineering
coming
in
and
learning
about
how
to
deploy.
Machine
learning
model
is
exactly
possible
because
I
think
businesses
will
want
to
put
features
that
can
they
like
they
can
get
from
machine
learning
into
their
products
as
soon
as
possible,
and
for
the
principal
of
agile
that
we
have
weaver,
obviously
likes
a
solution
that
would
remove
the
barriers
and
the
hardness
that
it
is
to
make
a
machining
model.
B
This
doesn't
have
to
be
the
soda
right,
the
standard,
the
state-of-the-art
kind
of
model
that
we
talked
about
in
research,
your
kind
of
your
academics,
but
it
can
be
a
simple
model
which,
for
example,
has
access
to
seven
percent
accuracy.
I
think
most
businesses
will
be
okay
with
that
and
they
would
like
software
engineers
to
come
in
and
and
then
help
them
easily
deploy
and
manage
this
kind
of
machine
learning
models
in
their
applications.
B
So
they
can
provide
some
kind
of
systems
right,
because
you
often
have
I've
like
been
through
some
machine
learning
courses,
and
we
often
have
the
examples
of
recommendation
systems
where
you
can
recommend
a
user
something.
But
what
bothered
me
about?
Those
was
that
when
I
was
sitting
down,
I
shoot
and
implement
kind
of
that
kind
of
thing
in
jupyter
notebooks.
I
realized
I
was
given
a
data
that
was
already
collected
from
somewhere
and
while
I
could
give
some
protection
like
I,
I
can
give
you
some
accuracy
or
position
over
there.
B
B
A
Now,
what's
what's
supported
in
in
the
main
cloud
environments
is
based
on
the
demands
from
customers
for
particular
types
of
solution.
So
it's
a
bit
of
a
circular
situation.
Right
now
in
the
the
cloud
environments
are
providing
jupiter
notebooks
in
production,
because
that's
what
the
customers
are
asking
for,
and
so
there's
there's
been
a
drift
towards
that
particular
way
of
working,
but
there
hasn't
been
sufficient
work
to
evaluate
whether
that
was
the
right
direction
or
whether
it's
going
to
lead
to
escalating
problems
in
the
future.
A
Hence
why
we
set
up
the
the
road
map
itself
so
that
we
could
paint
the
bigger
picture,
because,
if
you
think
about
it
most
of
the
work
that's
going
on
in
machine
learning
at
the
moment
is
small
teams
who
are
experimenting
to
see
if
they
can
build
models
that
will
improve
their
existing
products
or
allow
them
to
release
new
products.
A
So
they've
got
minimal
practical
experience
and
they're
they're
just
trying
to
do
whatever
is
easiest
to
run
an
experiment
and
and
try
and
incorporate
it
into
an
application
yeah.
So
that
means
that
most
people
are
at
the
at
the
bottom
of
the
ladder.
In
terms
of
the
journey
to
having
a
machine
learning
application
in
production
at
scale
for
an
extended
period
of
time,.
B
A
So
so
they're
not
yet
aware
of
all
of
the
problems
that
are
likely
to
be
encountered
as
they
climb
up
that
ladder.
So
lots
of
people
have
made
lots
of
approximations.
A
So
what
there's
there's
a
there's?
A
large
number
of
people
trying
to
enter
at
the
bottom
of
the
ladder
and
statistically
about
80
percent
of
them,
are
failing
to
make
it
into
production,
because
the
the
bits
that
they're
doing
only
get
them
so
far,
and
then
they
run
into
a
show
stopper
problem
and
they
they
can't
go,
live.
A
So
so
what
we've
been
trying
to
do
is
work
back
the
other
way,
rather
than
try
and
create
new
technologies
for
ml
ops
as
a
standalone
practice.
B
A
Within
jenkins
x,
you
can
treat
machine
learning
as
a
just
an
another
asset
in
your
build
pipeline.
A
A
You
can
use
a
quick
start
for
a
machine
learning
project
and
you
can
select
a
a
particular
machine
learning
platform
and
a
basic
algorithm,
and
it
will
generate
a
template
project
for
you
with
with
all
of
that
set
up.
B
Right
so
for
today
maybe
having
someone
else
and
what
should
we
talk
about
today
in
today's
meeting.
A
So
the
the
thing
that's
significant
from
our
perspective
is
that
we
need
to
make
sure
that
we've
got
all
of
the
updates
in
for
the
2021
document.
A
So
it'd
be
really
helpful
if
you
were
able
to
find
half
an
hour
at
some
point
to
to
read
through
the
the
mlops
roadmap
as
it
stands
today
and
then
let
me
know
if
you've
seen
any
patterns
or
challenges
that
that
we
haven't
included
in
in
last
year's
document
and
then
we'll
get
an
update
in
for
for
that.
B
So
you
mean
this
thing
right.
I
found
it
from
github.
B
B
So
I
will
go
through
this
and
I
will
read
more
about
it
and
if
there's
anything
addition
that,
I
think
can
be
done
I'll,
let
you
know
or
I'll
make
a
podcast
if
that
kind
of
works
as
well.
A
Okay,
that's
brilliant
right!
Well,
thank
you
for
for
coming
along.
It's
been
interesting
chatting
with
you
and
look
forward
to
to
having
you
along
in
the
future
and
having
some
contributions
on
the
document.
B
Thank
you.
That's
that's.
Awesome
yeah,
like
I'm
super,
excited
to
be
like
a
part
of
something
like
this,
because
I
feel
like
some
like
envelopes
is
something
that
I
think
will
blow
up
in
the
future
as
machine
learning
has
done
before,
but
instead
of
like
the
r
and
d
part
of
it
we're
moving
towards
the
production
kind
of
part
of
it
right
and
how
we
can
quickly
appropriate
machine
learning,
changes
in
the
model
and
stuff
like
that.
B
So
I
think
mlaps
is
the
right
kind
of
perspective
on
that
kind
of
thing
and
there's
something
that
I
can
unlock.
I
actually
want
to
work
professionally
sometime,
if
I
can,
as
a
machine
learning
engineer.
So
I
guess
a
bit
more
background
on
me
was
with
them.
I
started
learning
on
machine
learning
back
in
last
year's
june
and
july
much
more
seriously.
I
got
a
certification
and
and
indeed
running
from
deep
learning.ai.
B
But
after
that
I
felt
like
I
shall
think
about
what
are
some
ideas
or
points
that
I
can
make
that
have
machine
learning
in
them
right
and
I
was
quickly
dumb
struck
because
I
could
not
find
any
easy
way
of
putting
my
models
out
there
or
I
was
very
limited
to
what
technology
I
was
using.
So
I
have
like.
I
only
knew
one
way
of
doing
this
in
this
kind
of
thing,
but
I
did
not
even
know
if
this
came
in
this
other
technology
right,
because
I
could
not
find
a
drill
on
it.
B
B
So
I
think
that's
an
interesting
journey
but
yeah
I
I'll,
be
sure
to
contribute
and
ask
you
if
I
have
any
questions
that
would
be
pretty
good
as
well,
so
I've
ever
left
with
that
sure
for.
A
Sure,
brilliant
all
right!
Well,
it's
it's
been
good
to
chat
to
you
and
I
look
forward
to
hearing
your
feedback
on
the
document.