►
From YouTube: CDF - SIG MLOps Meeting 2021-06-03
Description
For more Continuous Delivery Foundation content, check out our blog: https://cd.foundation/blog/
A
I'm
actually
in
jersey
in
the
channel
islands.
Oh.
B
A
Yeah
well
certainly,
over
the
past
year
it's
been
quite
nice
to
be
stuck
on
an
island
in
the
middle
of
nowhere.
B
A
So
I
don't
know
how
many
people
would
like
to
see
today.
This
is
usually
a
fairly
quiet
session.
A
I'm
interested
in
trying
to
progress
things
a
bit
further
in
in
terms
of
getting
the
road
map
up
to
date
and
also
making
sure
that
we're
we're
talking
to
the
right
people
and
and
getting
the
right
requirements
in
place,
because
I
can
see
that
there's
there's
a
bit
of
a
gap
at
the
moment
between
the
work.
That's
going
on,
building
sort
of
devops,
tooling
and
ml
ops,
tooling,
versus
the
the
works
that's
going
on
in
in
building.
A
Yes,
the
supporting
infrastructure
and
platforms
to
to
actually
run
a
lot
of
the
machine
learning
work
on,
and
it
would
be
nice
if
we
could
join
some
of
those
dots
together
so
so
that
we're
getting
a
you
know
a
clearer
picture
on
both
sides
about.
You
know
what
capabilities
are
needed
out
of
the
the
pipeline
tooling,
and
you
know
what
capabilities
are
available
in
the
in
the
platforms
and
how
we
can
solve
some
of
the
shared
challenges.
A
From
from
your
perspective,
what
what?
What
would
your
thoughts
be
in
in
that.
B
Area
so
yeah
I
had
a
read
through
with
the
the
road
map
as
it
stands,
so
I'm
just
scrolling
down
to
kind
of
where
the
gaps
were.
You
see
it's
at
the
bottom.
A
A
So
you
you've
got
the
the
data
science
contributors,
who
are
very
strong
in
the
in
in
the
mathematics
and
are
working
hands-on
at
building
models.
A
But
they
have
limited
exposure
to
the
physical
hardware
and
limited
experience
with
managing
what
are
effectively
software
assets
in
production
environments.
A
Then
you've
got
the
the
people
who
are
building
ml,
ops
and
devops,
tooling,
and
obviously
they've
got
very
strong
cloud
backgrounds
and
lots
of
experience
in
managing
software
assets,
but
very
little
experience
in
machine
learning
and
again
very
little
exposure
to
specialized
machine
learning
hardware,
and
then
you've
got
the
hardware
providers,
who
are
obviously
right
on
the
cutting
edge
of
what's
possible
in
terms
of
accelerating
the
machine
learning,
but
also
identifying
challenges
that
come
with
trying
to
do.
Machine
learning
at
you
know
a
cluster
scale.
A
A
So
it
would
be
nice
if
we
could
make
some
more
connections
between
some
of
these
groups
and
then
try
and
spell
out
what
the
big
challenges
are
right
now,
because
you
know,
I
think
we've
got
a
reasonable
understanding
of
a
lot
of
the
problems
between
the
machine
learning
space
and
the
the
pipeline
management
space
in
terms
of
deploying
assets.
A
But
I
don't
think
we've
got
such
a
good
understanding
of
how
that
needs
to
scale
to
work.
Well
with
the
you
know,
with
the
dedicated
hardware
that's
available,
so
I
think,
there's
you
know,
there's
an
understanding
of
how
you
might
use
existing
public
cloud
infrastructure
to
to
do
this
stuff,
but
I'm
not
sure
that
we're
covering
off
all
the
bases
when
it
comes
to
you
know
using
a
dedicated
cluster
of
gpus
to
to
to
work
on
very
large
scale
problems.
B
Yeah,
I
guess
there's
the
the
training
aspect
so
sort
of
yeah
training
on
on
a
cluster
multi-node
training
and
then
there's,
I
guess,
the
inference
side
of
things
in
terms
of
hardware.
You
know
so
acceleration
of
of
the
inference.
So
I
guess
yeah
I
it
sounds
like
it's
more,
the
the
former
it's
the
disconnect
yeah
between
you
know,
training,
petabytes
of
data
on
a
huge
cluster
of
nodes,
so
so
yeah.
I
think,
maybe
what
I'm,
what
I'm
missing
is
kind
of
the
like
a
bit
of
a
background.
B
I
kind
of
came
here,
probably
not
knowing
exactly
what
we're
trying
to
achieve.
Maybe
that's
where
I'm,
where
I'm
kind
of
liking
lacking
some
knowledge.
So
are
we?
Are
there
people
sort
of
in
the
community
that
are
you
know,
working
on
or
developing
basically
solutions
to
the
problems
that
the
community
has
identified
here
or
what?
What's
the
the
goal.
B
A
Sorry
yeah,
so
so.
The
two,
the
two
areas
that
are
where
we're
we're
really
trying
to
provide
value,
is
one
to
communicate
an
overall
picture
of
what
all
the
challenges
are
in
the
hemalopse
space
to
everyone:
who's
working
in
that
space.
A
So
about
making
sure
that
there's
one
central
point
where
you
can
get
a
picture
of
all
the
challenges,
not
just
the
ones
that
are
related
to
the
things
that
you
might
be
working
on
locally
and
then
the
the
the
other
idea
is
to
try
and
bring
together
some
of
the
the
people
who
are
working
on
tooling
and
get
them
to.
A
B
A
And
that
is
not
the
case
with
the
mlop
space.
We've
we've
almost
gone
backwards
a
decade
in
in
terms
of
people
building.
You
know
very
bespoke
solutions
that
are
very
non-standard,
there's
very
little.
Interoperability.
A
In
terms
of
what's
going
on,
specifically
within
the
cdf,
we
have
several
projects
that
are
working
on
different
aspects
of
these
problems.
So
the
tech
on
itself
is,
is
one
of
the
cdf
hosted
projects?
A
So
so
that's
that's
become
if
you
like
the
de
facto
standard
for
running
pipelines
within
kubernetes,
so
so
we're
trying
to
put
as
much
of
the
standardization
as
we
can
into
into
tecton,
so
that
all
of
the
other
projects
that
are
based
on
tecton
then
inherit
that.