►
From YouTube: CDF SIG MLOps Meeting 2020-05-07
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
So
I
invited
a
few
other
people
what
they
might
join
up
a
bit
later.
So
there
we
go
rule
he's
from
cloudBees
and
it's
worked
in
our
various
data
science
things.
So
what?
If
you
view,
other
people
would
show
up
this
week,
Terry
so
maybe
before
we
start
introduce
themselves
of
it
so
I
know
all
of
you
better
Terry
tell
us
until
tell
Jesse
and
roll
stuff
you've
been
working
on
and
about
you.
Yeah.
C
B
Hey
I
think
invite
so
I'm,
just
no
mic
through
startup
community,
so
I'm
founder
of
code
lingo.
So
we
were
interested
in
basically
getting
insights
out
of
source
code
and
I
just
have
a
general
interest
in
ways
to
improve
the
productivity
of
developers
and
development
teams.
We
don't
actively
use
ml
or
AI
in
code
lingo
at
this
stage,
but
we
certainly
have
in
our
future
roadmap
and
yeah
super
engine.
B
It's
just
really
helpful
to
get
a
bit
of
clarity
around
this
definition,
so
it's
lovely
to
be
at
least
a
fly
on
the
wall
in
a
think
group
like
this,
because
I
think
it
is
I
really
love
a
lot
of
the
points
that
was
made
in
that
readme
and
it's
just
lovely
to
get
a
bit
of
clarity
around
the
terminology
and
and
in
the
problem.
Space
yeah.
Oh.
A
A
The
CDF
sir,
for
for
Jesse
enrolls
benefit
is
a
sub-sub
foundation
of
the
Linux
Foundation,
so
the
sort
of
big
in
the
open-source
community
and
it's
sort
of
a
very,
very
small
sister
of
the
CN
CF
and
other
other
things
like
that.
Roy
old.
You
want
to
mention
some
of
interesting
stuff
that
you
encountered
in
the
past
that
sounded
like
you
are
pretty
interested
in
the
I.
D
Started
with
matching
that
mean
some
years
ago,
at
Otto's,
working
on
the
first
implementation
of
the
service
service,
so
that
was
the
first
times
of
a
loop.
No,
yet
his
parka
streaming
was
not
ready
and
basically,
as
a
developer,
I'd
fall
in
love
with
this
kind
of
new
challenges
that
machine
learning
took
for
me
to
to
build
something
useful
and
after
that,
some
years
ago,
I
started
on
cookies,
where
I
am
working
on
the
more
DevOps
side
of
things,
TBT
team,
so
I'm
very
involved
in
releases.
D
That
kind
of
things,
and
now,
when
Michael
told
us
about
these
I
thought
like.
Oh
and
I
can
get
a
past
amazing
thing:
a
marriage
with
Kareem
amazing
thing
and
it,
and
it
also
makes
a
lot
of
sense.
So
it
was
like,
and
my
interest
come
from
from
the
icy
era,
sharing
a
lot
of
missing
things
that
can
be
done,
and
it's
quite
interesting
for
me.
So
that's
the
reason
why
I'm
here,
thanks
Mike,
for
the
invite.
A
Yeah
so
I
got
interested
in
this
through
talking
to
Terry
and
I
guess
my
day-to-day
interest
is
in
applying
some
of
this
stuff,
so
I'm
a
bit
of
a
I
think
the
word
is
dilettante
was
with
things
like
this,
like
I,
dabble
in
things
and
I've
used
different
libraries
and
tools
throughout
the
years,
but
lately
I
got
my
hands
on
a
lot
of
duck.
We
started
collecting
a
lot
of
data
for
jira
issue
tickets
and
you
know
sort
of
give
up,
see
data.
So
I
started
looking
at
things
that
we
can
do
by
analyzing.
A
It's
not
so
much
about
the
what
you
can
do
with
ml
or
I,
it's
more
about
how
you
get
to
production
and
auto
production.
A
lot
of
stuff
around
pipelines
in
that
so
I
think
that's
pretty
much
an
accepted
definition,
but
part
of
the
idea
of
sort
of
this
special
interest
group
is
to
narrow
down
sort
of
that
definition.
A
bit
because,
like
Terry's
mentioned,
there's,
there's
interest
from
all
sorts
of
interesting
companies
like
I
think
was
the
Intel.
A
C
There's
the
with
machine
learning
applications
in
production
environments
in
the
majority
of
cases
they
are
having
to
treat
those
things
as
special
cases
within
the
overall
software
development
lifecycle,
they're,
often
having
to
build
their
own
platforms
to
allow
them
to
deploy
those
assets
and,
as
a
result,
there's
this.
If
there's
a
lot
of
spoke
technology
out
there
right
now,
which
doesn't
sit
cleanly
with
the
sort
of
the
rest
of
the
learning
experience
that
we've
got
from.
You
know
70
odd
years
of
software
development,
experiments
and
learnings
snow.
C
Pants
were
a
lot
of
knowledge
that
we've
gained
from
that
software
space
back
into
the
machine
learning
teams
who
who
typically
tend
to
come
from
a
maths
background.
Rather
than
a
software
engineering
background,
and
in
many
cases,
then
there
have
been
exposed
to
the
types
of
challenges
that
we've
spent
least
daily
in
trying
to
run
software
in
production
environments.
C
A
That's
what
we
sort
of
went
over
last
week.
There
was
a
in
that
document.
There
was
a
table
towards
the
bottom,
and
the
first
thing
we
talked
about
was
sort
of
the
role
of
notebooks,
so
I
think
this
week
we're
going
to
move
down
that
list
and
talk
about
handling
data
and
stuff
I
had
a
few
other
things.
I
wanted
to
just
bring
up,
though
I
thought
word
would
be
interesting
just
quickly.
A
So
when
was
this
library
that
tensorflow
has
I've
mentioned
it
to
Terry
in
an
email
called
a
didn't
it
just
paste
it
in
the
chat.
If
you
can
see
that
I
came
across
that,
because
I
was
using
a
service
that
used
it
under
the
covers
and
I
was
curious
about
it.
So
I
started
looking
into
it
a
bit
more
and
I
thought
this
is
like
Terry
talks
about
you've
got
data,
scientists
that
are
whipping
out
our
code
or
now
code
and
notebooks
and
and
then
you've
got
engineer,
is
trying
to
bring
that
to
production.
A
But
this
is
almost
the
inverse
or
the
into
that
yang.
You
might
have
developers
that,
in
my
case,
I
have
access
to
a
lot
of
data
I'm,
not
a
data
scientist
I
can
do
the
training
courses
and
I
knew
a
lot
of
that
mess
once
that
I
find
tools
like
this
interesting
because
it's
like
it
removes
me
from
twiddling
or
too
many
knobs
with
the
I
prefer
edits
and
things
like
that
or
even
choosing.
You
know
how
many
layers
and
things
like
that.
A
It's
a
tool
that
will
it
uses
machine
learning
to
select
a
reasonable
approximation
or
not
approximation,
but
which
model
to
use
and
how
to
tune
it
and
I've
had
some
pretty
good
early
results
for
that.
So
I
thought
that
was
an
interesting
tool
and
I
mentioned
it
to
Terry
is
something
we
could
add
one
day
to
sort
of
a
library
of
quick
starts
so
that
when
a
developer
comes
across
this,
they
can
go.
I
can
format
my
data
this
way
and
I
can
normalize
things
that
way.
A
But
I
don't
really
want
to
learn
about
all
this
stuff.
We
don't
have
data
scientists
on
staff,
so
I
thought
that
was
an
interesting
tool
and
something
that
Google
are
investing
in.
So
it's
called
aid
in
it
and
it
uses
a
neural
network
to
train
other
neural
networks
and
even
our
ensemble
like
one
ensemble,
then
you
got
bunch
of
different
algorithms
or
different
types
of
neural
networks,
sort
of
glommed
together
and
that's
something
that's
usually
hard
to
do.
Just
cuz,
the
sheer
permutations
you
can
try,
and
so
this
works.
A
If
you've
got
a
good
like
you
might
not,
you
might
not
have
data
science
as
much
data
science
experience,
but
if
you
can
throw
computers
at
the
problem
in
massively
parallel
scale-
and
you
can
have
it
try
a
bunch
of
permutations
for
you
and
I
thought
was
fascinating
idea
and
with
with
sort
of
you
know
the
cloud
having
lots
of
stuff
on
top.
It's
starting
I
think
it
might
take
off.
So
so
that
was
that
was
an
interesting
one.
I
want
to
bring
up.
B
So
what
I'm
thinking
of
just
putting
my
traditional
software
development
head
on
we've
got
the
old
sdlc,
but
it
seems
like
there's
a
similar
life
cycle
that
needs
to
be
acknowledged
when
we're
looking
at
integrating
models
into
into
but
action
services,
and
that's
everything
from
you
know
your
initial
idea
to
actually
building
the
model
tweaking
the
model
and
then
actually
and
then
deploying
that
as
well
and
I.
Imagine
there's
a
whole
set
of
tools
and
and
mental
models
to
help
us
grapple
each
one
of
those
stages.
C
C
C
B
Cd
has
a
reason
as
a
concept
and
CI
has
erosive
as
a
concept
and,
for
example,
code
lingo,
where
we're
trying
to
get
across
this
idea
of
continuous
integration,
which
is
a
which
is
another
story,
but
I'm
I'm
wondering
if
there's
some
and
of
course
stop
me
here
if
I'm
railroading,
but
if,
if
it's,
if
there's
any
work
being
done
on
again,
some
kind
of
equivalent
to
an
S
DLC
to
help
guide
this
practice.
Let's
put
it
like
that,
like
what's
the,
but
what.
C
A
I
guess
Jesse's
sort
of
asking
like
how
do
you
like
if
you
use
one
of
these
things,
how
do
you
step
back
and
see
the
big
picture
of
like
what?
What
is
what
is
train?
What's
a
training
stage
of
a
pipeline?
What's
the
input
to
the
training's
where,
where
you
know,
if
you
were
to
draw
a
left-to-right
sort
of
you,
know,
you'd
like
to
do
that,
where
do
the
data
scientists
live?
And
you
know
things
like
that
exist
I.
Think
sorry.
So,.
C
We
need
to
demonstrate
that
there
are
there
ways
of
adding
value
to
the
machine
learning
ways
of
working
that
exist
to
date
without
constraining
those
systems
too
much,
because
a
part
of
the
challenge
is
that
we,
we
can't
say
here's
the
best
way
of
working
in
the
machine
learning
space,
because
it's
it's
a
technology,
that's
in
its
infancy
and
and
we
we
have
no
understanding
of
how
it
will
actually
turn
out
that
the
techniques
are
evolving
so
fast
that
it's
very
difficult
for
any
one
person
to
keep
abreast
I.
Think.
A
Terry
I
think
Jesse
comes
from
probably
a
similar
position
to
me
where
it's
sort
of
the
developer
angle,
as
here
are
the
things
I
know
what
at
what
are
the
things
I
don't
know,
whereas
the
more
sort
of
pressing
problem
is
what
Terry's
saying,
which
is
you
know
from
the
notebooks
and
data
science
angle,
because
that
stuff's
already
happening,
but
maybe
maybe
like
just
a
dictionary
somewhere,
would
help
like
a
dictionary
definitions
like
training
is
like
this.
Its
inputs
are
these
its
output.
Is
that
a
pipeline
for
machine
learning
senses?
A
Is
this
a
you
know,
a
model
is
a
deployable.
You
know
what
even
is
a
bottle
like.
That's
that's
something
that
a
lot
of
people
in
practic,
probably
even
diverse,
learn
just
don't
really
know
undercover.
It's
like
it's
like
well,
the
tensorflow
model
is
a
bunch
of
config
files
and
binary
protobuf
file,
like
you
know,
there's
meat,
maybe
there's
a
space
for
some
definitions
like
that,
but
that
would
I
kind
of
just
kick
that
in
my
head,
I,
just
sort
of
look
things
up,
I
think.
A
Think
for
for
developers
for
sure,
like
you
know,
model
models
could
be
big.
They
might
go
on
a
phone,
they
might
go
on
a
service
you're.
You
there's
a
lot
of
code
that
goes
into
massaging
data
before
it
goes
into
a
training
stage,
and
you
want
to
you
know
just
when
the
training
up
puts
a
new
model.
You
want
to
do
the
CD
kind
of
rollout
thing
with
that,
just
as
if
it's
you
know
a
jar
file
from
a
Java
compiler,
there's
there's
a
lot
of
things
like
that.
C
A
C
A
So
it's
it's
in
there's
a
change
pending,
for
you
know
a
pull
request
to
change
that
so
I
guess
we
could
move
on
so.
C
A
C
A
So
you
wrote
that
Kuban
not
know
jupiter
notebooks
are
the
tool
used
to
train
data
scientists
as
they
can
easily
be
used
to
explore,
like
don't
data
scientists
use
jupiter
notebooks,
not
the
other
way
around.
So.
A
C
D
I
believe
we
are
dealing
with
here
sure
is
I
have
felt
as
impedance
mismatch
when
a
data
scientist
is
talking
with
a
developer,
so
the
feeling
I
have
when
you
were
discussing
this
thing
is
we
are
having
that
exact
problem?
It's
like
we're
talking
to
that.
A
scientist
are
we
talking
to
developers?
Are
we
talking
to
both,
and
this
connects
to
what
you
was
talking
before
I
believe
so
maybe
some
sort
of
impedance
match
between
terms
vocabulary
needs
and
that
you
know
things
may
make
this
more
readable
for
both
sides.
D
C
Yeah
I
think
you're
you're
right
in
that,
but
I
think
we
also
have
another
problem
here,
which
is
just
the
habitual
one.
So,
typically
the
the
universities
are
using
a
limited
set
of
tools
to
teach
machine
learning
and
they're
doing
it
from
an
academic
perspective.
So
what
they
teach
is
what
you
need
to
do
to
get
an
academic
understanding
of
machine
learning,
but
they
don't
teach
any
of
the
practical
elements
of
how
you
use
those
assets
in
in
the
real
world.
A
C
C
C
Let
me
change
that
to
educate
and
then
that
will
that
will
be
clearer
yep
so
because
we
saw
the
same
problem
with
the
University.
So
initially
everyone
was
was
being
taught
to
program
in
C
and
and
then
we
had
to
retrain
everyone
when
they
came
out
into
the
industry
and
then
the
university
switched
over
to
Java
and
then
everybody
gets
trained
to
be
a
Java
developer
and
then
you've
got
to
retrain
them
again
when
they
come
out
into
industry.
B
C
C
You
know
the
reality
is
the
Jupiter
books
are
very
inconsistent.
You
can
you
can
execute
part
of
a
program
in
the
Jupiter
notebook
and
get
the
result,
which
is
complete
nonsense
because
you
only
executed
part
of
of
your
notebook,
but
it
will
still
give
you
a
result
and
it
that
result
is
meaningless.
C
So
so
the
the
point
here
is
really
that
what
we
need
to
do
is
flag
the
fact
that
other
tools
are
needed
and
encourage
people
who
are
working
in
in
in
those
spaces
to
develop
the
capabilities
to
support
the
ml
op
standard
with
with
their
tooling
and
obviously
the
the
likelihood
is
that
we
were.
We
are
going
to
see
things
like
Visual
Studio,
in
introducing
new
functionality
that
is
better
aligned
to
supporting
data
scientists,
and,
at
that
point
probably
tubes
and
notebooks
will
be
go
back
to
being
a
more
niche
desktop
right.
Yeah.
A
A
You
do
because
I
just
do
it
in
code
and
put
comments
in
there,
and
you
can't
remember
why
you
did
it,
whereas
when
you
flip
it
around
in
where
the
comments
comments
are
rich
in
first
class
and
and
the
code
is
little,
paragraphs
like
literate
programming
is
then
you've
got
this
big
trail
of
why
you've
made
that
decision
to
drop
that
column
or
why
you
normalized
it
in
a
certain
way.
Is
that
is
a
lot
of
machine.
A
D
C
A
So
one
I
guess
they're
looking
at
the
next
thing
down.
In
fact,
the
next
few
things
I
think
are
all
related
like
treat
MLS
it's
it's
first-class
citizens,
you
know,
develops
process
providing
mechanisms
by
which
training
sets
training
scripts
and
so
does
Rapids
may
L
be
versioned
how
their
auditable
across
their
life
cycle,
training
sets,
is
managed.
Assets
to
me
sort
of
the
overarching
answer
to
jump
to
a
solution
is
the
what
people
call
get
ops,
I
guess,
like
every
every
I'm
sure,
probably
agree.
A
A
C
C
So
we
mostly
fleshed
out
this.
This
section
I
think
we
have
a
fairly
broad
coverage
of
a
lot
of
the
key
problems
in
this
space
right
now.
So
what
we're
doing
from
this
point
on
is
working
our
way
through
each
of
these.
These
challenges,
which
which
we've
already
spelled
out
and
and
then
looking
at
the
technology
requirements
that
are
brought
up
by
those
challenges,
to
try
and
spell
out
how
we
might
develop
new
capabilities
to
address
those
challenges.
C
So
so,
what
we're
doing
is
collaboratively
fleshing
out
this
document
to
increasing
levels
of
detail
so
that
we're
we're
telling
the
story
of
our
thinking
process
is,
which,
which
can
then
be,
you
know,
shared
across
the
whole
community.
So
anyone
interested
in
building
a
product
in
this
face
has
got
a
very
firm
foundation
to
to
work
from
and
also
where
we're
looking
to
build,
and
it's
based
methods
for
solving
some
of
these
problems.
C
A
C
C
A
D
B
It's
the
impedance
between
the
software
developer
and
the
data
scientist
just
terminology
or
is
a
also
a
philosophical
difference
there,
and
it
would
be
useful,
slash
appropriate
to
have
a
crack
at
addressing
that,
because
I'd
certainly
be
quite
interested
to
see.
Terry.
How
you
see
the
world
from
your
perspective
and
I
could
certainly
give
it
a
crack
kind
of
how
I
could
see
the
world
from
from
from
a
software
developers
perspective.
B
C
B
Is
an
example
could
be
like
one
potential
difference
where
we're
talking
earlier
about.
You
know
the
work
works
on
my
machine
experience
with
the
jupiter
notebooks
and
that
kind
of
and
of
one
proof
point
I
feel
it's
quite
a
common.
It's
almost
a
religion
amongst
software
development
developers
that
the
code
doesn't
work
unless
it's
tested.
You
know
so
it's
kind
of
really
drilled
into
us,
then
the
importance
of
unit
tests
and
integration
tests,
and
just
because
it
works
for
you
there.
C
C
C
C
So
so
that's
that's
one
of
the
fields
of
of
testing
that
we
do
need
to
consider,
and
we
also
have
a
lot
of
awareness
around
things
like
bias
and
fairness,
where
we
need
to
be
able
to
establish
that
the
decisions
that
models
are
making
are
not.
You
know,
inherently
biased
in
some
way,
and
they
act
with
reasonable
levels
of
fairness
across
the
the
populations
that
the
model
is
intended
to
serve.
C
So
so
we
are
introducing
a
new
level
of
testing
complexity
into
the
into
the
problem
space,
which
introduces
a
new
set
of
challenges
in
terms
of
how
we
actually
test
for
that.
So
so,
partly
this,
this
question
does
need
to
discuss
how
we,
how
we
treat
machine
learning
assets
in
respect
of
things
like
unit
testing
and
integration
testing,
but
also
how
we
extend
those
paradigms
to
cope
with
things
like
detecting
ethical
problems
or
baiance.
C
A
One
of
the
reason
this
stuff
sort
of
comes
up-
it's
quite
topical
in
the
news
like
people
will
say.
Well,
how
does
the
model
explain
how
it
made
a
decision
which
is
never
easy
to
do
because
they're
often
a
black
box,
it's
not,
but
that
doesn't
necessarily
mean
it's
impossible.
It's
it's
the
same.
If
you
wrote,
if
you
wrote
a
whole
lot
of,
you
know
messy
cohesive
statements
and
loops
and
stuff,
and
then
it
makes
a
decision
to
approve
a
loan
or
not
how
do
Jessie's
a
lot
of
things?
You
know
what
I
mean
yeah.
A
All
real
production
systems
are
built
under
duress
that
way,
but
the
reason
is
you
can
trace
back
through
the
changes
in
the
code
and
who
approved
the
request.
Was
it
covered
by
testing?
Did
the
testing
cover
these
scenarios?
Did
it
have
acceptance
testing?
And
you
know
it's
kind
of
the
same
thing?
If
you
can
do
all
that
with
the
model,
it's
not
like,
you
have
to
crack
open.
We
have
many
layers
of
new
runs.
It
is
and
doesn't
make
sense,
it's
more
like
you
know
what
set
of
data
trained
it.
Can
we
reproduce
it.
C
A
A
If
it's
a
recommendation
system-
or
you
know
a
movie
thing
or
it's
gonna
suggest
you
know
in
just
this
case-
it
might
be
suggesting
if
it's
for
some
code-
or
you
know
a
policy
change
or
something
like
that-
that
it's
spotted
you,
you
would
work
better.
This
way,
there's
not
really
an
issue
there.
It's
like
it's
just
a
recommendation
or
a
bit
of
automation.
It's
not!
But
in
you
know
the
grand
scheme
of
things,
it's
super
risky
to
not
be
able
to
explain
things
and
trace
it
back.
A
C
C
So
so
we're
already
seeing
a
strong
skew
against
being
able
to
use
machine
learning
solutions
in
certain
territories,
and
that
means
that
the
standards
that
and
any
AI
solution
will
be
held
to
will
be
much
higher
than
the
equivalent
standard
for
a
human
worker
in
the
same
situation,
and
so
the
expectation
for
AI
will
in
many
cases
be
unrealistic
in
terms
of
compliance,
because
we
have
no
mechanisms
by
which
we
could.
We
could
prove
the
same
thing
with
the
existing
humans
doing
doing
the
role.
C
B
B
C
A
C
My
my
consideration
in
that
space
is
that
host
organization
are
going
to
be
required
to
prove
legislative
compliance
in
there
nice
Aleutians,
and
they
will
want
to
include
that
in
the
governance
processes
for
their
sort
of
releases.
So
so
it
really
needs
to
be
a
feature
of
CI
CD
platforms
going
forwards
that
there
are
mechanisms
built
in
that.
Allow
you
to
to
demonstrate
having
gone
through
that
audit
process.