►
From YouTube: Kubernetes Machine Learning WG 20180412
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
Okay,
so
the
agenda
is
empty,
so
maybe
you
can
just
have
a
quick
meeting
today.
I
had
one
thing
that
I
wanted
to
discuss,
which
is
like
at
a
high
level.
There's
lots
of
existing
ecosystems
that
are
trying
to
do
some
sort
of
like
ml
solution,
with
or
without
kuvira's
I've
been
having
this
question
on
like
how
can
we
be
effective
and
actually
make
a
difference
in
this
space
by
us?
I
mean
like
folks
who
are
participating
in
as
I
was
working
room?
A
Maybe
just
a
kick
kick
start
this
the
start
process
I
was
thinking
that
our
goal
could
be
I.
Think
I
think
Connor
had
summarized
this
pretty
much
in
the
in
the
umbrella
issue.
That
haired
file,
which
is
our
goal,
is
to
make
it
easy
for
for
both
power
users,
as
well
as
like,
as
far
as
like
SAS,
like
solutions
as
in
as
in
use
cases
that
have
some
amount
of
scale,
and
in
this
case,
like
specifically,
are
on
ml.
To
succeed
with
Coronas
is.
Is
that
fair,
like
good?
Would
that
be?
A
If,
as
a
group,
if
we
can
just
pick
one
problem
that
we
all
think
is
like
really
really
important
and
I'm
just
like
make
sure
that
works
really
well
and
then
move
on
to
the
next
or
I
mean
we
could
do
multiple
things
to
buy
like.
If
we
can
have
like
some
concrete,
achievable
goals,
I
feel
like,
then
we
can
structure
a
conversation
just
around
there
and
then
start
forming
some
execution
plans.
B
C
B
D
Yeah
so
currently,
some
of
our
internal
users
are
using
PVC
and
then
some
folks
in
ku
flow
community
also
are
trying
to
use
it
and-
and
we
have
interest
from
so
pachyderm
project,
a
pachyderm
is
another
workflow
and
pipelining
framework
for
running
ml
workloads,
so
packet
on
was
trying
to
integrate
with
goo
flow
and
they
are
looking
at
kvcs
the
glue
for
data.
So
that's
another,
you
user
story
that
is
emerging
for
kvc
use
case
or
at
least
the
data
set
management.
But
but
in
terms
of
what
KBC
salts,
it's
it's,
it's
kind
of
one.
D
D
There
is
some
background
noise,
but
this
is
going
to
repeat
what
you
said
if
I'm
not
wrong,
so
you
you
said,
as
a
group,
we
should
concentrate.
One
of
the
goals
you
could
concentrate
is
improving
the
data
set
management
issues
or
problems
that
are
being
faced
by
the
users
within
the
kubernetes
ecosystem.
Is
that
correct,
yeah.
A
Okay,
I
personally
of
the
data
set
idea
in
that,
like
storage,
consuming
storage
right
now,
is
this
a
problem
with
curators
and
and
it's
something
that
ties
a
lot
into
the
existing
cumulus
architecture
itself,
and
so
I?
Don't
see
it
as
like
an
app
that
you
can
deploy
on
top
of
cube
because
it
needs
some
underpinnings
within
Q
at
least
yeah.
A
A
So
if
you
can
like
identify
certain
user
journeys
today,
where
people
suffer
and
then
like
then
come
up
with
what
we
could
do
differently
there,
then
that
would
help
or
if
there
are
specific
use,
cases
that
are
already
known.
Those
can
be
shot
here
that
will
help
justify
or
actually
understand
how
we
can
make
a
difference
in
that
area.
D
Okay,
but
why
why
that
templating
both
take
both
Connor
and
I,
meant
about
how
we
do?
How
do
we
get
from
co2
container
to
kubernetes
at
least
that's
the
templating
idea,
so
right
right.
So
so
probably
we
can
make
that
clear
and
and-
and
that
seems
to
be
a
little
e
obvious
problem-
more
obvious
issue
with
almost
all
data
centers
in
ml
ml
practitioners,
where
they
good
sorry
so.
A
B
So
it's
it's
I!
Guess
it's
in
some
ways
kind
of
similar
to
what
cube
flow
is
trying
to
do,
but
with
the
connection
that
it
also
can
help
you
build
the
container
it
it's
released,
kind
of
as
a
as
an
initial
alpha,
just
in
the
concept
of
like
release
early
and
get
feedback,
I
think
as
far
as
like
how
the
the
architecture
works
and
how
the
how
the
code
gets
deployed.
B
A
If
we
can,
if
we
can
make
whatever
we
build
genrich's,
it
could
apply
for
any
different
class
of
problems
and
that
it
just
becomes
a.
If
you
can
imagine
a
plug-in
model
like
transfer
is
just
a
plugin,
then
that
that
same
construct
could
apply
for
any
other
source
image
problem
that
exists
in
the
Cuban
airspace.
A
C
C
A
E
Once
a
cache
we're
actually
having
a
conversation
with
scaffolding,
next
Monday,
so
you
could,
you
guys,
could
come
to
see
gaps
on
next
Monday
April
16th
and
ask
Apple
team
they're,
gonna,
demo,
scaffold
and
they're
very
interested
in
building
containers,
but
I'm
not
entirely
sure
whether
they're
interested
well.
We're
gonna
talk
about
that
in
general,
yeah
kind
of
like
what
that
looks
like.
A
A
A
A
I
think,
having
a
having
a
simple
proposal
in
the
m/l
working
group
in
the
criminals
community
repository
like
stating
what
are
the
pain
points
and
like
what
could
be
the
use
of
Chinese
I.
Think
that
that
that
would
be
a
great
starting
point,
because
we
can
say
the
storage
stick,
for
example,
and
have
them.
B
D
D
A
D
Yeah,
so
there
are
three
categories:
I
think
in
this
one
is
people
some
scientists
and
data
vendors
and
are
like
comfortable
with
CLI.
The
other
category
is
UI.
Third
category
is
they
want
to
just
work
you
from
the
core
itself?
I
think
there
are
I
mean
I
mean
I,
can
only
think
of
those
three
right
now,
but
there
might
be
more,
but
but
each
of
these
solutions
are
different.
Probably
so
maybe
we
want
to
think
think
over
that
as
well.
That
could
be
I.
A
D
Yes,
and
and
especially
like
looking
at
these
three
different
categories,
so,
for
example,
in
graphical
UI
case,
it's
a
Jupiter
hub
in
code.
First,
there
is
like
solutions
like
meta
particle
in
CLS
case.
There
are
several
solutions,
including
case
on
that
pin
cool
flow
Q
flow
and
there
might
be
other
solutions
out
there.
Yeah
right.
A
D
A
A
F
Yeah
we
mostly
in
cupola,
we
provide
you,
know
very,
very
low
hanging,
fruits
and
syntactic
sugar,
so
that
we
provide
a
a
form
form
that
sort
of
allows
us
to
easily
spawns
notebook
images
that
are
that
we
maintain
and
curate
that
are
there
a
good
set
for
you
know
ml
related
workflows
right.
So
it's
very
low-level.
Very
you
know
not
much.
You
know
it's
not
we're
not
adding
a
whole
lot
on
Jupiter
hubs.
A
Well,
Jim
I,
don't
mean
to
like
I,
don't
mean
to
stop
us
from
investing
that,
but
if
you,
if
you
have
some
ideas
on
how
to
preserve
itself,
could
be
extended
more
in
order
to
like
integrate
better
with
with
cumulus
I
think
those
are
useful
things
to
explore
as
well.
I
think
anyone
who's
not
here,
has
some
ideas.
Last
year
on
on
further
improving
integration
with
Jupiter
have.
F
Yeah
I
think
the
the
the
biggest
question
that
I've
come
up
with
looking
at
with
looking
at
Super
Hub
is
that
to
some
extent
they're
they're
building
with
a
platform
which,
for
like
data
scientists,
which
kind
of
makes
sense
because
they
they
don't.
They
don't
necessarily
want
to
assume
you're
running
on
kubernetes,
and
so
some
of
the
things
that
they
do
to
support
non
burn
areas.
F
F
I've
talked
to
you,
you
me
before
and
I
think
what
he
told
me
basically
was
that
you
know
where
they
see
it.
I
think
uptake
of
Jupiter
hub
in
kubernetes
is
people
trying
to
scale
sort
of
massively
so
either
because
of
like
you
have
I
think
like
a
common
use,
cases
like
massively
mock,
MOOCs
or
mocks
whatever
those
the
the
appropriate
pronunciation
is
where
you
want
to
have
sort
of
a
whole
bunch
of
people
in
the
class
using
the
same
Jupiter,
Hoglund
and
jib,
and
making
that
scale
out.
That's
where
kubernetes
then
comes
into.
E
A
If
that
is
a
project
that
this
working
group
would
be
interested
in
and
I
can
drive
it
through.
This
working
group
I
would
love
to
get
support
on
Windows
and
I.
My
my
expertise
would
be
limited
to
Linux
as
far
as
I
know
like
Mac
cannot
handle
external
GPU.
So
if,
if
I
ever
left,
my
own
means
I
would
probably
just
add
support
on
Linux,
but
I
said
as
if
as
a
community,
we
can
collaborate
there
and
I
get
it
working
on
those,
and
maybe
on
that
would
be
awesome.
A
A
That
works
well.
G
A
E
A
A
Less
cost
effective,
it's
more
about
yeah,
it's
just.
Everyone
has
a
laptop
as
a
developer,
so
it's
like.
Okay
just
have
to
buy
this
external
device
again
I'm
hypothesizing
here
I,
don't
have
concrete
yeah.
If
to
me,
it's
like
lowest
in
the
priority
like
water,
walk
like
what
I
would
like
to
see
is
like
mini
cube.
Working
for
simple
use
case
is
like
a
Linux
laptop
or
like
an
Alienware
laptop
or
for
sure
definitely.
A
For
people
I
use
right
are
like
the
dgx
workstation
is
that
that
that
fief,
some
people
buy
so
I
mean
that
could
keep
extending
that
could
be
like
more
hardware
in
the
future
and
like
I,
think
this
would
extend
to
like
a
six
and
FPGAs
to
so.
If,
if
there
is
enough
interest
in
that
area,
then
that
could
be
another
project
that
we
can
drive
and
I'm
happy
to
share
that
yeah.
G
Sing,
that's
I,
seen
a
story,
interesting
problem,
I'm
working
in
Emacs
of
research
and
we
are
lab
of
sixty
scientists
and
the
speaker
walks
for
you,
you
see
is
that
everyone
has
a
workstation.
So
it's
not
a
laptop,
but
you
see
Lisa
machine
with
local
GPU.
Why
you
run
the
first
test
just
to
me
to
make
sure
Sally
see
that
your
model
is
achlys
works
when
you
scale
written
on
a
cluster.
That
will
be
something
like
flow
because
you
don't
want
to
GPUs
for
ten
years.
D
E
E
D
E
A
A
B
I
guess
it's
related
to
a
proof
of
concept
we
have
going
internally,
and
maybe
it
doesn't
make
sense
for
it
to
cross
into
this
this
worker,
but
basically
it's
a
way
to
just
put
a
little
bit
more
structure
around
the
concept
of
an
experiment
for
a
data
scientist,
and
so
they
can
have
like
some
some
result.
Metadata,
maybe
pointers
to
you,
know
their
output
directory
for
each
job
run
on
shared
storage
or
blob
storage.
A
A
F
There's
a
lot
of
activity
around
this
already
you
sort
of
in
the
in
the
community
right
so
like
you
know,
there
is
a
cart,
cart
tip
or
is
it
cat
some
cat
tip,
which
is
basically
a
busy
a
busy
a
clone
that
somebody
basically
just
introduced
into
coop
flow
and
is
hosting
it
under
coop
Club,
and
it's
built
on
model
DB
to
provide
both
sort
of
hyper
parameter
tuning
and
then
using
model
TVB
for
providing.
You
know,
model
browsing
and
experimentation,
and
then
tender
board
as
well
is
also
sort
of
looking
at.
You
know
this.
F
This
problem
of
trying
to
you
know,
surface
people
theta
and
make
it
easier
to
surface,
and
you
also
have
other
projects
like
Studio
ml
that
are
working
in
this
space.
I
think.
But
my
view
is
that
some
of
these
projects
might
benefit
from
you
know
using
crts
under
the
hood.
But
to
me
they
sort
of
seem
like
apps
in
and
of
themselves
that
we
would
run
on
kubernetes
to
provide
this
functionality.
D
So
I
want
to
mention
here
that
I
mean
when,
when
things
work,
it's
all
finest
in
like
when
then
things
work
as
they're
expected
to
work.
It's
all
fine,
but
the
problem
comes
when
when
we
want
to
when
you're
doing
some
hyper
hyper
parameter,
optimization
and
something
goes
wrong
or
what
you
are
doing,
trying
to
do
some
it
when
something
goes
wrong,
let
it
do
the
jobs
and
the
logs
of
the
jobs.
A
F
B
There
are
a
couple
of
I
mean
yeah.
It's
definitely
like
for
the
volume
that
you
know
results
could
generate.
You
want
like
a
more
scalable
data
store,
but
in
general,
like
you
know,
for
experiments,
if
you
were
to
keep
track
of
that
as
a
CR
D,
for
example,
there's
usually
like
at
least
one
order
of
magnitude,
less
experiments
than
jobs,
and
so
the
scalability
of
that
doesn't
seem
to
be
a
concern.
F
To
me,
the
the
advantage
of
the
of
the
control
of
the
controller
pattern
is
that
you're
sort
of
managing
some
set
of
infrastructure.
You
have
multiple
resources
that
go
through
some
states
and
you
have
to
manage
those
resources
throughout
their
lifetime.
Keep
them
up,
keep
them
healthy
right,
and
so,
when,
once
you
get
passed
to
your
your,
your
job
is
finished
and
your
or
your
experiment
is
done,
and
you
just
have
some
record
of
that
experiment
that
you
want
to
persist
and
it's
immutable.
B
A
B
A
I
just
want
to
make
sure
you're
not
confusing
two
issues.
One
is
like
just
handling
storage.
In
that
every
time
you
launch
a
new
new
job
or
new
experiment.
Do
you
need
to
like
provision
storage,
and
that
could
be
just
like
scratch,
space
that
needs
to
be
persistent
or
it
could
be
like
the
actual
model
data
that
you're
publishing?
A
Lots
to
I,
don't
deny
that
you
could,
in
theory
like
have
your
own
connectors
and
like
have
data
bases
and
stuff
like
POSIX
file
systems
or
object,
stores
like
I,
mean
or
textures,
are
not
even
represented
by
storage,
API
CL,
which
is
a
separate
problem.
But
you
could
go
that
down
the
route
of
using
databases,
but
just
sort
of
be
on
your
own.
At
the
point,
I
mean
this
by
itself
is
not
gonna
help
you
it's
not
going
to
improve
your
life
anyway.
There.
A
That
was
one
on
storage
and
the
second
one
is
the
archival
and
I
think
Jeremy
brought
up
good
points,
which
is
there's
a
scalability
limit
sort
of
implicitly
in
that,
like
there's
a
scalability
limits
on
how
much
CRTs
we
can
have.
So
we
have
to
like
be
limit.
How
many
actual
objects
we
can
store.
We
can
throw
against
API
server.
As
far
as
I
can
tell.
There
is
no
archival
solution
available
for
criminals.
Isis
know
that
it's
like
not
even
a
single
database
that
you
can,
that
has
been
like
designed.
A
A
F
We
would,
we
would
be
I
view
it
in
a
generic
problem
and
we
would
like
to
see
it
solve
the
generic
problems,
but
we
have
a
proposal
floating
around
in
coop
cloud
to
solve
this
and
basically
it
follows
the
same
sort
of
pattern
as
cluster
level
logins,
where
you
basically
just
have
a
Damien
in
your
cluster
that
monitors
the
API
server
and
then
emits
the
objects
to.
You
know
some
back-end
that
it's
that
it's
configured
to
talk
to
right.
So
it's
very
pluggable,
very
customizable,
I,
think
I.
F
Think
one
of
the
questions
that's
kind
of
floating
around
is
whether
we
should
just
admit
the
things
as
text
and
sort
of
assume
that
you
can
use
whatever
logging
back
in
that
you're
currently
using
or
whether
we,
whether
a
columnar
datastore,
would
be
more
appropriate
for
some
of
the
it's
data,
provenance
and
ETL
sorts
of
analyses
that
we
think
we'd
want
to
run
to
support.
So
we
want
so
like
the
data
provenance
question.
A
Okay,
this
thanks
for
that
update,
Jeremy,
I
I
still
feel
like
it's.
A
the
problem
of
our
table
is
like
a
gendering
resolves
like
a
generic
one,
and
it's
probably
much
easier
to
talk
about
rather
like
trying
to
do
it
specifically
for
certain
classes
of
apps
within
cutest.
I
also
feel
like
every
every
increment,
as
provider
would
probably
have
their
own
choice
of
yeah.
F
So
so
I
agree
with
you
100%
and
I.
Don't
think,
like
anything
that
we
would
actually
do
or
proposed
is
actually
specific
to
Kubler,
I'm
now
and
so
I
think
we
would
love
to
see
that
it
gets
up
streamed
and
solved
in
kubernetes,
like
this
notion
that
you
can
automatically
persist
an
archive
of
all
of
your
records
and
integrate
that
with
like
audit
logging
and
data
provenance.
Logging
like
that'd,
be
fantastic,
but
you
know
it's
mostly
about
speed
of
execution
and
easier
to
produce
hype,
and
then
we
can
always
upstream
later
yeah.
A
A
A
F
So
in
that
sense,
I
see
it
is
as
a
problem,
and
so
so
as
one
concrete
example
with
Argo,
we
use
our
NGO
for
CI
CD
in
coop
flow
and
we're
finding
that
we
submit
so
many
jobs
that
eventually
slows
down
things
like
in
this
case.
It's
the
it's
the
UI,
it's
not
the
which
I
think
is
related
to
the
API
server
performance,
but
it
we
end
up
having
to
delete
these
things.
So
it
is
an
issue.
Okay,.
A
A
A
Okay,
I
want
to
make
sure
that
all
the
points
you
mentioned
are
actually
received
and
digested.
You
also
mention
metrics,
but
metrics
is
something
that
is
going
to
be
very,
very
close
to
each
of
those
apps
that
are
running
right
in
this
case
each
of
those
ml
frameworks.
So
I
don't
know
what
cumulus
can
do
in
a
generic
manner
and
that,
like
there's
already
four
medias
integration
and
there's
like,
and
it
should
be
pretty
easy
to
expose
metrics
as
from
meteors
data.
A
B
You're,
actually,
you
know
kind
of
like
progressive
values
that
let
the
data
scientists
know
how
training
is
progressing
so
things
like
loss
or
some
specialized
f-1
metric,
or
something
like
that.
So
yeah
I
think
you're
right.
It's
totally
external
there's,
just
more
of
illustrating
how
it
would
be
used.
Yeah.
A
D
A
A
Yeah
I
think
part
of
the
problem
is
that
it
who
were
authors,
such
a
guy,
has
to
have
some
good
understanding
of
the
frameworks
themselves
and
finding
those
people
in
the
cumulus
community
is
sort
of
hard
because
they're,
mostly
dealing
with
with
like
Gendry,
curious
concepts
and
so
I.
Don't
know
if
I'll
be
successful
in
finding
a
person
to
victim.
A
A
A
If
we
can
so,
we
identified
a
bunch
of
things
today,
source
image,
tear
sets
mini
cube
and-
and
there
was
also
like
now-
she
does
really
cube
and
then
there's
stateful
jobs
and
storage
management
for
each
and
every
job
run.
And
then
there
was
archival
monitoring,
I.
Think
if
you
can
at
least
start
getting
deeper
on
any
one
of
these
topics.
That
would
be
awesome.
A
D
Sure
I
can
at
least
start
southern
issue
and
and
discuss
why
we
created
the
KBC
project.
Maybe
that
will
be
a
good
starting
point
and
we
can
add
more
more
details
to
that
later,
on
I
mean,
but
but
but
I
think
some
of
it
was
already
presented,
but
we
can
go
into
more
details
of
each
of
those.
Each
of
those
points.
Ok,.
D
A
A
A
The
other
thing
we
could
do
is
like
once
we
we
mean
a
month
or
so
from
now,
once
the
drill
down
a
little
bit
more
on
the
topics
that
we
have
identified,
we
could
probably
go
and
present,
and
the
criminis
community
meeting
after
the
the
mean
kunis
community
meeting
and
sort
of
give
everyone
a
heads
up
that
hey.
We
are
you're,
organizing
ourselves
through
this
forum
and.
A
G
D
G
A
lot
about,
what's
the
scale,
that
trying
to
reach
with
this
project
and
see
whether
we
think
this
column
rank
is
right.
What
would
be
missing
to
run
something
adults
on
kubernetes?
The
paper
is
pretty
far
out,
so
let
me
give
you
the
link.
They
insist,
it
might
be
awful
games
that
is
and
finding
what's
where
it
will
break.