►
From YouTube: CDF SIG MLOps Meeting 2020-05-21b
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
C
A
B
A
A
B
D
B
B
B
B
The
other
piece
of
work-
that's
going
on
at
the
moment,
is
adding
a
basic
glossary
of
terms
so
that
we've
got
some
standard
definitions
for
the
various
pieces
that
we're
discussing
the
the
the
session.
This
week
was
primarily
talking
about
the
the
differences
between
the
standard,
C,
ICD
and
DevOps
behaviors,
and
the
more
specific
and
unique
aspects
that
we
we
need
to
think
about
from
a
machine
learning
perspective.
So
specifically,
you
looking
at
the
scaling
challenge
of
how
do
you?
B
How
do
you
work
with
very
large
numbers
of
either
GPU
or
TPU
accelerators
in
a
machine
learning
perspective?
How
do
you,
how
do
you
go
from
a
pure
kubernetes
view
of
the
world
which
is
cpu-based
to
to
a
model
in
which
you
may
need
to
be
targeting
dedicated
hardware
and
then
also
discussing
some
of
the
challenges
around
how
we
manage
the
data
flow
component
within
machine
learning
world?
B
A
A
Implementations
for
these
narator,
you
are,
you
have
certain.
You
have
certain
technologies
in
mind
which
you
are
looking
at.
Was
this
something
you
are
asking
the
coming
to
develop
around
and
and
in
general
the
other
part
of
it?
Is
you
know
some
something
like
you
know,
machine
learning,
models
consuming
GPUs
etc?
Is
this
I
think
this
is
definitely
you
know
something
which
is
from
a
larger
develops
perspective?
It
makes
a
lot
of
sense
right.
B
B
The
blue
areas
are
areas
where
we
we
know
we
need
stuff
and
stuff
is
being
built
and
the
yellow
areas.
Are
they
the
areas
where
we
we
actually
have
big
unknowns,
where
this
potentially
no
proper
activity
going
on?
So
the
the
view
of
the
roadmap
is
that
what
we
need
to
do
is
is
take,
though,
both
a
medium
term
and
a
long
term
view
about
what
what
capabilities
we
actually
need
within
ml
ops
and
then
to
paint
the
picture
of
you
know
what
what's
already
there.
What
can
we
do
today?
B
You
know
what
can
we
do
with
cube,
though?
What
can
we
do
with?
You
know
the
components
that
already
exist
and
then
I
like
what
gaps
there
are
and
where
we
need
to
evolve
solutions
in
the
future
to
meet
those
requirements.
So
so
this
is
about
giving
everyone
a
picture
of
where,
where
there
are
opportunities
in
the
space,
where
existing
solutions
or
new
solutions
can
can
be
developed
to
to
fill
those
gaps
and
and
fit
a
need.
A
A
There
is
a
quite
a
bit
of
work
happening
there
and
has
already
happened
there
around.
You
know
trying
your
models,
which
are
being
deployed
to
GPUs
or
CPUs,
are
choosing
what
hardware
you
want
right.
What
sort
of
capacity
from
that
hardware
you
want
and
most
of
these
projects
under
the
KF
long
rail
are
now
publishing
roadmap
dot,
MD
right,
which
is
essentially
more
tactical,
I,
would
say
a
year.
Look
at
what
we
think
is.
A
You
know
the
direction
for
the
project
for
the
year,
so
it
might
be
worthwhile
looking
and
checking
as
well
from
that
perspective
and
I
think
you
know
there,
this
can
help.
Is
you
know
there
are
like
projects
under
the
Q
flow
around?
You
know,
distributed
training,
hyper,
parameter,
optimization
models
serving
metadata
pipelines
right,
but
beyond
that,
if
there
are
other
emerging
areas
right
where
the
community
should
be
looking
at
right.
So
obviously
you
know
things
like
federated
learning.
A
Right
or
when
you
start
being
going,
you
know
bit
of
layers
on
top
right,
auto
ml,
Auto,
a
right
and
then
they're
still
not
something
right
which
which
on
which
you
know
there
are
projects
which
we
have
spawned.
Another
queue
flow
umbrella,
but
I
think
you
know
there
they
are
getting
quite
a
bit
of
traction.
We
have
seen
a
lot
of
requests
now
coming
around
federated
learning
or
so
to
say
you
know,
multi-cloud
learning
different
people
are,
you
are
giving
it
different
names
where
or
ng
learning
we
are.
A
B
So
I
I'm
seeing
a
lot
of
interest
in
the
wearable
space,
for
example,
where
we
expect
scenes
to
see
a
lot
of
target
device
is
needing
inference
on
on
edge
devices
in
wearable,
which
will
have
to
be
very
small,
dedicated
hardware,
so
so
that
drives
the
discussion
about
cross-compilation
and
and
delivery
into
into
a
six.
If
you
like,
and
then
I'm,
also
seeing
a
lot
of
customer
demand
in
the
industrial
space.
So
there's
there's
a
lot
of
work
going
on
with
sort
of
industrial
Internet
of
Things
and
factory
automation.
B
Stuff,
like
that,
where
there
are
some
some
very
interesting
challenges,
especially
around
the
latency
of
systems
and
the
need
to
be
able
to
train
against
very
large
data
sets
but
inference
in
real
time
very
close
to
machines.
That
are,
you
know,
either
safety
related
or
are
dealing
with
very
high-value
processes
where
it's
critical,
that
they
can.
They
can
make
complex
decisions
at
high
speed.
B
No,
so
you
know
a
big
reason
for
having
the
roadmap
really
is
to
start
to
paint
the
picture
of
all
of
these
quite
complex
and
re
O's
in,
in
which
a
basic
cloud-based
solution
is
it's
not
gonna,
be
fit
for
purpose
and
that
we
need
to.
You
need
to
actually
be
thinking
ahead
and
designing
much
more
subtle
approaches
to
how
we
manage
some
of
these
problems,
but
still
bring
this
overarching.
A
A
C
Vias
Bracy
are
currently
looking
into
case
serving
and
the
design
for
like
models
like
current
score
or
discount
customer
lifetime
value
and
putting
our
models
into
production
through
care
serving
and
trying
out
to
build
API
so
automatically.
So
this
is
really
particularly
interesting
for
us,
and
also
the
scalability
of
cube
flow
is
very,
very
handy,
and
it
helps
us
keep
costs
on
like
decent
amounts.
A
A
This
one
right,
so
this
was
a
talk
at
cube,
cone
right
at
which
myself
in
play,
if
he
is
the
CTO
of
Selden
right,
we
came
last
year.
This
was
in
one
of
the
details
and
also,
you
know,
has
links
to
the
slack
channel,
etc.
But
you
know
how
old
escapes
are
doing
work
that
was
essentially,
you
know
the
focus
I
mean.
Essentially,
if
you
see
the
the
companies
who
came
together
to
build
this
right,
I
mean
Selden
is
started
from
London
right
and
and
they've
created.
Selden
models
are
being
engine
in
open
source
right.
A
The
rest
of
the
names
you're
probably
more
familiar
with
right,
but
the
idea
was
to
do
this.
You
know
on
on,
you
know,
scale
to
zero
model,
because
the
point
of
you
know
model
serving
engines
taking
all
the
models
deployed.
Taking
a
lot
of
the
GPU
resources
are
still
sitting
and
not
being
able
to.
You
know,
scale
down
to
zero.
If
there
is
no
traffic
right,
it
has
become
concerning.
So
that
was
one
of
the
primary
ideas.
A
A
A
You
need
to
respond
to
this,
but
I'm
having
more
there
right
so
and
so
is
you
know
the
other
folks
Dan
see
this
last
comment
which
spanned
over
to
47.
That
is,
you,
know,
Dan
Sony's
from
Bloomberg
right,
so
Bloomberg
is
totally
running
here,
serving
right
now
in
production
spread
across
no
multiple
node,
so
I
think,
if
is
probably
the
best
person
right
now
in
the
community.
Given
the
you
know,
the
amount
of
he's
he's.
A
A
A
You
know
and
that
you
know
sort
of
explains
the
flow
right.
How
how
the
flow
is
flowing
from
is
tio2
is
TM,
open
gateway
that
your
Fox,
you
know
it
it
has
depending
on
you
know
the
scenarios
you
are
getting.
Hopefully
you
know
and
one
of
the
resolutions
they
are
right,
but
the
goal
is
to
continue
growing
this
people.
Do
it
right.
The
decks
in
channel
general
has
been
challenging,
listen
escapes
because
Dex
Levin
will
allow
you
to
do
programmatic
authentication
and
you
need
small.
You
know
something
where
you
can
have
you
know
a
uie.
A
Actually,
you
know
generating
the
DWG
tokens
and
you
authenticating
and
authorizing
based
on
that,
but
the
latest
release
of
K
observing
right,
which
is
K,
observing
0.3
I,
don't
know
which
release
you
are
using
if
you're
using
0.3
that's
bought
along
with
Kennedy
of
0.11
dot.
2
is
what
we
are
recommending
to
get
around.
You
know,
so
he
attacks
probe
issues.
Yeah.
A
C
Ok,
yeah
I
mean,
if
you
read
more
about
these,
these
issues
not
sure
but
evolved
version
of
Kennedy.
If
we
do
have
it
first,
I
will
start
with
that
diagram.
I
think
it's
particularly
useful
that
we
showed
in
debugging
well
yeah
and
where,
for
example,
I
I
find
where
in
abuse
you
can
take
initiative
is
for
this
explained
ability.
Maybe
you
know
we
have
a
framework
for
called
responsible,
AI
or
explaining
our
neural
network
models
like
for
stability,
testing
and
bias,
and
other
kind
of
things
that
can
be
tested
on
neural
network
models.
A
C
A
Yeah,
because
we
have
like
so
just
to
give
you
an
idea
right
around
this
topic
in
general
trusted
here,
we
call
it
trusted
AI
right.
We
have
projects
around
a
fairness
which
is
the
AI
fairness
360.
There
is
adversarial
robustness
toolbox,
which
is
around
you,
know,
detecting
adversarial
attacks
and
on
defending
against
them
right.
This
is
being
used
right
now
by
DARPA,
for
you
know,
adversarial
attack
and
defense
mechanisms,
it's
funded
by
DARPA.
A
Yes,
let
me
share
right,
so
this
is
what
you
know.
This
is
if
you
go
to
just
get
a
Broadcom,
slash,
IBM,
you
will
see
right
so
the
AI
finest
360
is,
you
know
trained
here
all
right.
This
is
for
bias,
detection
and
mitigation
right
and
there's
a
pretty
cool.
You
know
website
also,
will
you
can
try?
It
out
run
some
demos
at
your
head,
etc
to
detect
and
mitigate
bias.
The
second
thing
is:
I
mentioned
the
versatile
robustness
toolbox
that
this
is
a
project
you
know
which
is
right
now
initiated
by
IBM.
A
Darpa
has
given
a
ground,
they
are
actually
using
it
in
defense,
research
right
so
pretty
popular
in
adopted
right.
So
a
lot
of
the
defense
mechanisms,
evasion,
attacks,
extraction,
attacks,
poisoning
attacks,
and
then
you
know
beyond
the
attacks.
How
do
you
actually
implement
defense
mechanisms
right?
So
if
you
are
interested
in
this
space,
there's
another
one
right
on
the
explain
ability
side
we
have
our
project
or.
A
Explain
ability
360
right
so
so
this
one
is
essentially,
you
know
quite
a
bit
of
algorithms
from
IBM
research
right,
and
that
also
has
its
website,
but
also,
you
know,
provides
an
interface
on
top
of
Liman
tap
right.
So
if
you
go
there
to
the
website
as
well,
wait
it's
the
state
in
the
github.
It
talks
about
you
all
the
different
methods
it
is
using.
You
know
you
can
run
your
own.
You
know
scenario,
demos
and
and
try
it
out
right.
What
does
it
mean
and
the
generating
explanations
for
different
kinds
of
consumers
so
yep?
A
A
A
C
A
A
He
explained
ability,
for
example,
right
when
you're
looking
at
we
just
you
know,
go
to
the
project
yourself.
This
has,
you
know,
explained
ability
at
different
levels.
So
if
you
can
expand
alright,
so
the
idea
is
that
you
know
what
we
call
you
know
a
there
is
explained
ability,
algorithms
for
the
data
itself
right,
but
when
you
actually
go
to
more
models,
you
know
we
have
two
areas.
One
is
you
know
local
explanation.
The
second
is
global
explanation
and
local
explanation
is
essentially
you
know.
The
model
results
right.
A
Global
explanation
is
very,
very
essentially
treating
model
as
a
black
box
right.
You,
you
don't
have
access
to
the
internals
of
the
model
or
the
model
is,
you
know,
fairly
complex,
deep
learning
model
right.
So
how
do
you
generate?
You
know,
explanations
for
this.
So
right
you
know
the
algorithms
one
of
the
algorithms
we
use
is
you
know
having
a
several
gate
model
right,
which
is
learning
over
the
explanations,
and
you
know
creating
its
own
modified
explanation
around
that
right.
A
A
F
F
Maki
Jackson
has
written
an
article
as
well
there's
a
bit
more
sort
of
generic
on
ml
ops
and
then
the
other
thing
I
was
going
to
ask
Terry
if,
like
I,
think
it
would
be
pretty
cool
to
stick
in
the
vision
from
the
roadmap
as
a
post
that
just
links
then
to
the
roadmap
again
just
another
way
to
get
some
eyes
on
it
and
increase
the
visibility
for
folks
to
see
so
Terry.
Would
you
be
okay
with
me
pushing
there?
I'd
put
it
under
your
name
as.
B
F
F
No
problem
and
in
general
I
think
we'd
like
to
start
using
things
like
the
newsletter
more
regularly.
So
we
have
a
recurring
theme
of
ml
ops,
so
maybe
every
I
don't
know
what
it
will
be
six
or
nine
months,
but
then
the
other
thing
we
want
to
tie
in
with
the
other
CDF
events.
So
if
folks
are
interested
in
either
participating
in
the
podcast
or
webinars,
please
ping
me.
Let
me
know
and
I'll
help,
sync
that
up
with
Jackie
who's
running
the
show
and
all
all
those
things.
Okay,.
A
Jc
one
question
I
have
is
like:
is
there
a
larger
for
the
overall
CD
foundation?
Right
I?
Think
one
of
the
things
we
do
want
to
now
do
is,
like
you
know,
get
the
folks
who
are
actually
behind
you
know
some
of
these
projects
like
Jenkins
or
you
know
things
like
Tecton,
etc,
be
a
little
bit
more
and
depending
right,
I
mean
even
with
that
audience
right.
A
F
So
what
I
was
thinking
there
and
that
the
general
body
where
we
have
a
lot
of
the
project
representatives
and
folks
from
the
communities
around
the
projects
is
the
the
technical
Oversight
Committee
I
know
towards
the
end
of
last
year.
It
was
kind
of
lost
a
bit
of
momentum,
but
Dan
Lawrence
has
come
in
as
the
new
chair,
so
he's
reviving
things
quite
well
and
what
he's
looking
to
do
is
get
updates
from
the
SIG's
on
a
regular
basis
of
the
interoperability.
Sig
did
an
update
at
the
last
talk.
F
A
Think
it's
probably
you
know,
I
mean
one
of
the
things
we
have
been
doing
as
I've
dedicated.
You
know,
activity
as
part
of
the
sig
was
you
know
the
project
around
this
right,
which
is
more
hands-on,
exercise
on
ground
exercise
will
be,
and
essentially
you
know
trying
to
build
something
and
deliver
something
right
which
can
be
used
right.
A
You
know
the
cue
flow
pipeline,
Python
interface
on
top
of
Techtron
right,
so
you
can
use
Python
programmatic
ways
of
defining
pipelines
on
pipelines,
so
that's
achieved.
The
second
phase
is
where
we
essentially,
instead
of
going
to
Tecton
directly,
we
want
to
orchestrate
from
pure
flow
pipeline
engine
and
the
reason
for
doing
that
would
be.
You
know
to
get
the
cue
flow
pipeline
UI
and
the
lineage
tracking
and
the
artifact
tracking,
all
all
integrated
around
tracfone
right,
so
I
think
you
know
I
would
love
to
figure
out.
You
know
how
to
now.
F
A
Both
its
primary,
you
know,
I
would
like
to
see.
You
know
there
are
folks
who
are
interested
in
evolving
right,
something
like
this
for
their
own
needs,
right
and
and
by
virtue
of
getting
users.
Also,
you
get
more
feedback
right,
like
you
know,
things
are
so
I
think,
even
if
they
don't
want
to
contribute,
but
if
they
have
a
need
right
where
they
are.
Essentially,
you
know
someone
who
are
running
Tecton
and-
and
they
want
to-
you-
know,
evolve
it
towards
machine
learning
needs
and
they
want
to
bring.
A
F
Okay,
this
sounds
like
something
we
can
connect
you
with
Jackie
to
do
a
webinar.
We've
now
got
CDF
running
regular,
webinars
and
I
know
she's
looking
for
topics
and
I
think
this
would
work
quite
well.
So
how
about
I'll
connect
the
two
of
you
and
I'll
point
you
at
the
the
forum
where
you
just
need
to
sort
of
book
your
slot
in
that
sound
good.
F
A
No
I
think
these
are
the
two
primary
I
mean
the
roadmap
activity
plus
you
know
all
hands
on
ground
technical
project
right
now
is
you
know
the
cue
flow
pipeline
and
check
on
integration,
which
is
happening
at
much
faster
speed
right
and
an
active
in
development.
The
end
today's
discussion,
as
you
saw
right,
I,
mean
depending
on
the
community,
needs
right.
So
today's
discussion
was
more
focused,
like
you
know,
around
model
serving,
and
you
know,
model
explaining
ability,
motive,
fairness
right,
so
that's
the
second
part
of
that
how's
that
you
know
the
pipelines.
A
Hopefully
you
are
helping
you
to
get
to
the
point
where
your
models
are
deployed
right,
but
how
you
deploy
that
in
them
in
production,
and
how
do
you
enable
things
like
explain,
ability?
First,
detection,
detection
right?
So
hopefully
you
know
we
start
getting.
You
know
so
more
focus
around
that
area,
and
then
we
can,
you
know,
think
of
carving
or
something
specific
in
that
space.
A
Thanks
thanks
a
lot
so
yet
Thomas,
it
will
be
great
to
you
know.
You
know,
get
some
more
details
around
your
responsibilities.
She
ate
it
and
see.
You
know
how
we
can.
There
is
a
trust
area
committee,
as
well
as
part
of
the
Linux
Foundation
AI
right,
which
essentially
is
more
focused
on
all
these
areas.
Bias,
explain
ability,
adversarial,
detection
and
generation
right.
So
if
you
are
interested
right,
you
can
reach
out
to
me
and
I
can
get
you
connected
enjoy
that
as
well.
If
that's
your
area
of
interest,
yeah.
D
I
have
two
questions.
This
is
Lois
also
from
PwC
another
engineer
and
the
same
team
as
Tomas.
D
A
This
right
so
essentially
the
integration.
Can
you
see
my
screen
yeah?
Yes,
the
integration
with
TF
exes
is
at
a
different
level
than
what
you
were
thinking
in
terms
of
the
Czech
term
right.
So
this
is,
for
example,
you
know
queue
flow
pipeline
and
it
works
with
our
goal
as
an
engineering
to
the
covers
right,
what
we
are
doing
is
replacing
our
goal.
Bringing
in
Tecton
all
right.
Rest
of
the
entry
point
remains
the
same
right.
So
essentially,
there
is
a
merge
happening
at
an
SDK
level,
so
the
first
version
of
that
SDK
is
ready.
B
A
A
You
know
either
the
same
SDK
to
draft
T,
FX
pipeline
or
a
kfb
pipeline
right.
So
that's
the
SDK
story.
We're
essentially
you
know
one
single
SDK.
If
you
define
a
pipeline
using
Q
flow
pipeline,
DSL,
semantics
or
using
you
know,
T
effects,
you
can
use
that
single
SDK
and
launch
it.
The
second
work,
which
is
a
bit
longer
term,
as
you
know,
aligning
around
the
same
DSL
as
well
right,
so
which
is
essentially
right.
Now.
A
Yes,
you
can
use
the
same
as
DK,
but
you
know:
there's
still
the
the
Python
DSL
for
both
are
quite
different
right,
so
how
you
will
define
a
T
FX
pipeline
versus
how
you
will
define
now
Q
flow
pipeline?
You
know
the
syntax
and
the
semantics
are
different
right.
So
that's
where
there
is
that
discussion
going
on
right,
heavily
and
and
part
of
it
is.
You
know
that's
something
which
the
internal
Google
teams
need
to
come
themselves
together
as
well
right,
which
is
the
tf-x
team
in
the
KP
team.
A
Presentation
right
so
assuming
you
have
a
line
around
SDK,
which
has
happened
right
so
hopefully
it
will
be.
You
know
more
widely
distributed.
Then
the
second
thing
was
DSM.
The
third
thing
is:
when
you
compile
you
want
to
compile
to
something
intermediate
right
and
not
specifically
to
Ann
Arbor
yeah
Melora
Techtron
y'know
by
virtue
of
compiling
to
something
intermediate
right.
A
There
is
now
an
opportunity
to
create
a
stern
industry
standard
right
around
how
pipeline
should
be
defined
in
a
more
neutral
language,
which
then
you
know
the
Chevy
engine
can
interpret
based
on
the
target
given
to
it
hey.
This
is
the
IR.
Your
target
is
checked
for
normal
or
tomorrow
in
future.
Maybe
it's
Jenkins
right
and
thou
shalt
be
able
to
convert
that
IR
back
to
that
corresponding
mean
yet
right.
So.
A
Phase
of
that
right,
which
is
you
know,
a
little
bit
more
more
in
the
future
right
so
long
answer
to
your
question
but
short
answer,
the
combined
SDK
is
ready.
The
next
will
be.
You
know,
looking
at
that,
we
have
the
same
DSL,
which
is
essentially,
you
define
using
the
same
way
whether
it's
at
the
FX
pipeline
or
KP,
and
the
third
angle
to
it
would
be.
You
know
you
also
compile
to
something
which
is
common
across
everything
and
the
engine
takes
that
common
standard.
A
D
Now
T
FX
also
kind
of
not
compiles,
but
like
you
can
also
use
it
as
a
structure
for
an
airflow
pipeline.
Is
it?
Is
there
a
possibility
that
we
can
use
that
intermediate
representation
structure
to
create
a
an
airflow
pipeline
that
you
can
around
perhaps
locally
on
your
machine
rather
than
and
grenades.
A
Ya
know,
I
think
the
tf-x
right
now
has
this
concept
of
runners
right.
So
airflow
is
one
of
the
runners
behind
he
effects
right.
So
by
virtue
of
it,
you
know,
assuming
you
the
there
is
a
single
SDK.
You
will
combine
compiling
to
a
single
IR
and
that's
that's
what
you
know.
Gfx
is
taking
right
and
then
you
know
what
is
the
runner
tf-x
has
chosen
right.
A
So
when
you
are
running
GFX,
you
have
chosen
airflow
as
a
runner
right
logically,
that
should
work
on
airflow
right,
but,
as
I
said
this
part
of
it,
where
there
is
this
common
IR,
which
we
all
are
using,
it's
still
I
would
say
at
this
point.
You
know
there
are
like
lot
of
e-mails
back
and
forth.
We
are
trying
to
establish
that
communication
and-
and
maybe
you
know
David-
this
is
also
an
opportunity,
an
area
where
you
can
help
where
I
think
you
know
we
are
heavily
interested
in
this
lift
has
come.
E
A
Know
so
I'm,
so
there
is
an
email
chain
right
between
how
I'm
going
on
with
Pavel
on
that
side
right,
it
would
be
I
mean
if,
like
with
what
wall
you
were
doing
with
ml
spec
right
or
the
pipeline
metadata,
or
you
know
finding
pipeline
inna,
so
I
think
I
would
see
that
you
have
a
inherent
interest
in
it
right
if
you
can
make
it
more
like
you
know,
also
that
yes,
Microsoft
is
heavily
interested
in
standardizing
this.
That
will
sort
of
speed
up
some
activity
on
this
side.
Right,
I
think.
E
A
A
Thanks
I
think
this
was
a
great
discussion
today
and
you
know
I
will
compile
and
send
an
email
with
the
links
etc.
Right
for
both
around
you
know
the
a
of
serving
trusted
AI
as
well
as
around
you
know,
the
tf-x
K,
if
we
work
right
and
Terry's
roadmap
is
in
the
seg
ml
of
3po
itself
right.
So
there
is
one
single
thing
to
go
to
that
right.
A
If
there
is
any
other
thing
right
which
you
need
in
terms
of
you
know
how
to
reach
out
on
different
areas,
whether
it's
the
trusted,
iPod,
biased,
explained
ability
or
the
checked
on
you
for
pipeline
integration
or
kfb
GFX,
you
know
just
should
open
email
and
I'll
be
well
yep
thanks.
Everyone
thanks
again
thank.