►
From YouTube: Kubernetes WG Batch Bi-Weekly Meeting for 20230622
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
A
Today
is
June
22nd,
my
name
is
mache
and
I'll
be
your
host
for
this
batch
working
group
meeting.
Today
we
have
a
couple
of
topics,
but
a
quick
announcement.
A
Kubernetes
AI
HPC
day
happening
during
cubecon
in
Chicago,
November
6
until
8.
If
I
remember
type
the
word
something
along
those
lines.
The
call
for
proposal
is
open
and
it's
due
on
August
6th.
So
you
roughly
have
a
month
and
a
half
to
submit
your
proposals
and
sign
up
for
for
the
AI
HPC
date
if
you're
interested
in
in
attending.
A
A
Hearing
none,
okay,
I'll
move
it
to
the
next.
In
that
case,
marching
do
you
want
to
share
your
screen.
C
A
A
C
A
Isn't
that
the
problem
where,
where
you're
sharing
the
screen,
you
can't
talk
I,
remember
that
they
fixed
it.
Okay,.
C
Okay,
there
is
another
button
that
was
moved
somewhere
else:
okay,
okay,
perfect.
So
today,
I
would
like
to
talk
about
the
Persian
request,
crd
that
I
have
AAP
proposal
in
in
this
link
here
and
yeah.
That's
that's
the
topic
I
want
to
cover
today,
so
so
the
motivation
currently
cluster
of
Skiller
does
not
provide
any
a
way
for
us
to
express
a
fact
that
group
of
PODS
would
like
to
have
a
capacity
provision
for
them.
C
Atomically
before
the
pots
are
created,
and
there
is
a
lot
of
situations
this
may
be
relevant
for
users.
A
couple
of
them
is
that,
for
example,
users
might
want
to
have
All
or
Nothing
semantics
for
space
scale
up
and
currently
due
to
like
interactions
between
workout
controller,
beat
like
job
controller
or
like
other
one
Cube,
scheduler
and
cluster
office
killer.
The
VMS
will
be
provisioned
in
in
a
step
function
anyway.
So
like
we
will
try
to
look
at
some.
C
The
VMS
are
there
and
are
idle,
basically
because
they
cannot
proceed
with
their
workloads
without
all
of
the
games,
and
there
is
also
like,
if
you
eat,
latency,
for
scale
UPS,
some
of
the
workloads
may
be
bursting
and
because,
because
of
the
same
reason,
users
will
actually
expect
bigger
latency
with
each
cluster
of
the
scalar
scale.
C
Upload
will
just
see
some
of
the
Bots
and
just
Spam
bunch
of
like
have
one
request
to
cloud
provider
to
create
some
VMS,
and
this
latency
will
be
like
bounded
by
the
slowest
VM
and
each
of
those
like.
C
Usually,
we
will
have
like
if
we
want
to
have
60
VMS.
This
will
take
not
six
times
the
request
of
getting
100
games,
so
there
is
some
some
latency
that
we
can
cut
off
there.
C
So
this
is
my
motivation
for
for
this,
and
we
are
proposing
like
an
API
which
will
we
call
provisioning
request
crd,
which
is
custom
resource
object
and
would
be
a
nice
namespace
object
where
users
can
describe.
The
group
of
the
pots
would
be
created
and.
C
Like
that,
they
want
to
have
a
capacity
for
for
them
provision
in
a
particular
Manner,
and
it
should
contain
like
the
template
of
the
pots
and
how
many
pots
there
are,
and
possibly
the
different
pot
templates
for
for
different
within
a
one
one,
provisional
request.
So
we
can
imagine
like
I
wanna
have
10
tops
like
10
parts
that
will
be
running
a
job
and
One
controller.
That
will
be
just
scheduling,
some
some
specific
tasks
to
them
and
it
should
allow
admins
to
like
support
different
implementation
of
provisioning,
the
capacity.
C
So
we
can.
We
can
try
to
have
Atomic
one.
We
can
try
to
have
like
other
modes
of
atomic,
so
maybe
the
step
function
but
make
it
configurable
in
a
way
that
I'm
happy
with
at
least
such
and
such
VMS.
A
note
not
the.
This
is
an
overview
of
of
how
you
would
how
the
young
of
the
one
of
the
provisioning
quest
would
look
like.
C
So
we
would
have
some
name.
Some
project
require
class
in
this
case,
like
the
atomic
scale
up
how
long
we
can
wait
for
the
VMS
to
to
brought
up
and
like
one
bot
set,
which
would
consist
of
20.
F
C
Yeah
and
the
question
so
far
I'm
here
none
so
here
we
have
like
generic
atomic
scale
up
and
within
the
within
the
EAP
that
I
have
I'm,
proposing,
also
a
genetic
check
capacity,
which
would
basically
be
one
of
track
to
verify
that
the
cluster
within
a
cluster.
There
is
enough
capacity
to
provision
a
given
set
of
spots.
C
This
is
not
a
real
guarantee
like
we
will
not
block
those
this
capacity
for
them
so
like
there
will
be
no
no
nothing
scheduled
on
down,
but
it
might
be
useful
in
scenarios
where
we
have
only
one
workout
emission
controller,
for
example
the
queue
in
in
the
scenario.
We
don't
expect
any
like
robots
to
hijack
this
capacity.
So
if
there
is
capacity
you
can
just
assume
in
the
IQ
Loop
that
it's
there
after
after
the
chart,
another
one
is
the
aforementioned
atomic
scale
up
and
yeah.
C
C
This,
for
example,
different
note
groups,
so
different
different
type
of
machines
and
whatnot,
and
if
we
succeed,
then
we
pass
this
information
to
user
that
if
we
fail,
we
move
all
of
the
partially
provisioned
VMS,
remove
them
and
like
retry,
some
like
like
later
in
some
explanation
back
of
manner
within
this
on
the
duration
that
users
can
specify,
as
mentioned
in
the
in
the
example
and
what
is
also
is
in
important.
C
Can
provide
a
specific
class
for
for
their,
like
Cloud
providers
can
provide
specific
classes.
So,
for
example,
if
one
of
the
cloud
providers
was
to
have
a
better
API
to
to
provision
VMS
in
a
way
that
we
can
go
there
and
say,
yeah
I
want
hundreds
of
those
VMS.
If
you
can
provision
them
provision
them
atomically.
C
If
don't
just
let
me
know,
I
I
will
retry
some
later
on
and
this
this
kind
of
logic
will
allow,
for
example,
users
to
to
not
pay
for
the
partition,
personal,
partially
provisioned
VMS
for
the
duration
before
we
delete
them
and
yeah
the
the
proposed
life
cycle
of
the
object
is
like
as
follows.
So
first
user
or
like
a
framework,
for
example,
the
queue
creates
the
object,
then
like
cluster
rotoscaler
picks
it
up
chooses
amount,
pools,
no
pull
or
one
and
tries
to
create
nodes,
or
this
is.
C
This
is
like,
for
example,
the
atomic
scale
up.
So
in
this
small
you
know
provisional
nodes
in
the
check
capacity,
we'll
just
check
whether,
if
it's
there
and
later
on,
we
pass
this
information
through
conditions
to
to
users,
whether
we're
successful
or
whether
we
are
retrying
or
whether
there
is
enough
capacity
or
not,
and
at
this
moment,
if,
for
example,
for
the
atomic
mode
users,
if,
if
you
are
successfully
created
the
VMS
users
can
create
Bots
and
users,
they
can
can
Mark
those
as
with
specific
annotation,
that
those
are
consuming.
C
This
specific
version
request
and
like
if
all
goes
well,
the
the
those
parts
should
be
scheduled
on
the
dedicated
capacity
that
was
provision
and
once
all
of
the
pods
are
scheduler
users
can
delete
the
provision
request.
Otherwise
it
will
be
like
garbage
collected,
yeah
I
have
I
sent
the
question
happy
to
answer
it.
As
of
now.
G
E
Yeah
thanks
for
presenting
so
I
had
a
question,
so
you
made
a
comment
that
users
don't
pay
for
partial
provision
VMS.
So
so
can
you
describe
that
a
little
bit
more.
C
Yeah,
so
the
one
of
the
modes
that
we
are
saying
is
the
generic
atomic
scale
up
and
in
this
mode
user
base
for
the
partial
Provisions,
because
this
is
like
we,
we
try
to
get
get
them
the
capacity,
and
if
we
don't
cannot
get
the
capacity
fully,
then
then
different
Cloud
providers
may
give
us
like
only
partial
provisions
and
in
this
one
users
will
file
for
the
capacity
depending
on
the
capacity
like
the
provider.
C
But
there
may
be
like
a
specific
cloud
provider,
API
to
atomically
scale
up
and
and
then
then
this
within
this
specific
class
we
might
we,
my
users,
might
not
be
required
to
pay
for
this
partial
provision,
because
those
will
not
happen.
Basically,.
E
It's
up
to
the
cloud
provider
how
they,
you
know,
give
you
the
resources
or
whatever
the
slas
of
that
Resource
Group
got
it
yeah.
C
Yeah,
like
Cloud
providers,
may
have
different
operation
class
like
and
those
might
have
different
guarantees
about.
The
atomicity.
B
One
comment
here
so
a
little
bit
explanation
how
a
cluster
Auto
scalar
is
organized,
so
we
have
an
interface
for
providing
resources
in
cluster
of
the
sky
that
is
implemented
currently
by
about
30
Cloud
providers.
We
will
add
something
new
to
the
interface.
However,
it
will
take
time
before
all
of
this
Verity
plus
Cloud
providers
respond
and
implement
the
API.
Some
of
them
may
not
have
this
capabilities
available
yet
so
as
a
stop
gas
solution,
we
want
to
provide
something
like
semi-atomic
like
this
generic
Scale
app.
B
That
gives
you
like
some
of
the
possible
Atomic
experience.
However,
in
generic
node
Cloud
specific
way,
sample
providers
may
have
something
better
already,
and
in
that
case
they
will
implement
this
API
differently,
like
use
some
different
ways
to
expand
the
number
of
nodes
in
the
cluster,
and
then
we
will
be
able
to
provide
better
guarantees
stronger
guarantees
that
once
the.
B
As
Google,
we
want
to
integrate
it
with
what
we
have
on
our
end.
So
the
idea
is
that
soon
there
will
be
Google
specific
implementation
and
generic
semi-atomic
implementation.
That
should
work
for
most
of
the
cloud
and
there
will
be
in
interface
for
other
Cloud
providers
to
follow
so
that
they
can
use
whatever
secret
API
or
public
IPS.
They
have
on
their
own
end.
G
Hi
can
can
you
hear
me
hey
thanks
Daniel
for
presenting
us.
We
actually
have
a
few
users
who
are
interested
in
this
type
of
game,
scheduling,
I,
guess
you
could
call
it
so
very,
very
interested
in
this
talk
going
forward
to
your
next
slide
about
the
the
way
it
works.
G
If
there
is
a
way,
maybe
with
a
controller
where
you
could
specify
the
pods
along
with
the
provisioning
request
and
then
once
the
capacity
is
available,
you
schedule
those
pods
from
moving
that
that
need
to
submit
the
pods
and
then
remove
the
provisioning
request.
If
that
makes
sense,
just.
G
C
Yeah
you
mean
like
fire
and
forget
mode
like
you
could
create
the
pots
and
the
provision
requests
at
the
same
time,
kinda
and
then
just
those
should
should
happen
along
the
capacity
is
provision
there
like
those
should
be
scheduled.
Well,
when
once
we
have
Bots
like
the
note
for
them,
yeah.
G
I
guess
some
abstraction
here
to
say
like
here's,
my
real
pod
spec
once
the
capacity
is
available
for
all
of
them
together
put
them
schedule
them
all
together.
B
That
would
be
problematic
because
that
would
require
additional
step
on
scheduler.
So
right
now,
in
the
picture
we've
got
only
cluster
Auto
scaler,
which
reacts
to
a
scheduling
decisions
right,
so
you
would
like
to
have
somehow
block
the
pods,
possibly
with
a
scheduling,
Gates
and
only
unblock
them
once
the
provisioning
request
is.
Is
there
right
something
like
that?
B
So
that's
that's
possible,
however,
possibly
as
a
responsibility
of
some
additional
controller.
B
So
you
create
that's
not
like
rocket.
Science,
probably
won't
happen
in
in
MVP,
it's
not
for
everyone,
but
yeah,
it's
something
that
could
possibly
be
it's
something
that
could
possibly
be
added
to
the
proposal
as
like
step
number
two.
G
D
If
I
may
answer
to
the
question
from
from
the
Q
perspective
this,
this
is
one
one
feature
that
we
can
easily
integrate
with
from
the
Q
controller.
If
you're
not
familiar
with
Q
controller
does
job
scheduling
and
it
supports
some
some
apis,
in
particular
the
job
API.
So
if
you
use
the
job
API
you
submit
a
job,
Q
can
create
a
provisioning
request
for
you
once
the
permission
is
ready,
Q
admits
the
job
and
all
the
ports
are
created.
D
That's
One,
One
controller
that
will
make
use
of
of
this
Pro
this
API,
making
it
seamless
for
the
researcher.
Let's
say.
B
Yeah,
that
is
the
plan
we
we
want
to
have
it
as
as
soon
as
possible,
and
we
want
to
make
this
API
kind
of
standard.
So
right
now
we
as
secure
the
skin
working
on
with
a
carpenter
folks
on
convergence
plans
between
these
two
autoscalers,
and
the
idea
is
that
this
API
will
not
only
be
supported
by
cluster
Auto
scalar,
but
also
by
contactor.
So
hopefully
it
becomes
kind
of
standard
way
of
asking
for
a
capacity
in
your
cluster.
F
A
I
have
a
question
related
with
the
API,
which
is
kind
of
related
to
what
Alex
was
asking
about
I'm
looking
at
the
provisioning
request
and
I.
Think
that
stands
out
is
that
I
need
to
copy
the
entire
pods
back
from
every
other
resource
that
I'm
working
with
can.
A
Can
we
make
the
API
a
little
bit
more
generic
such
that
we
will
either
provide
some
kind
of,
for
example,
pot
selectors,
so
that
it
knows
what
is
the
type
of
thought
and
it
can
read
that
part
from
the
cluster
and
get
that
information
from
the
cluster
rather
than
me,
copying
the
entire
prospects
from
my
job
that
I'll
be
creating
or
if
I'm,
going
to
be
using
it
for
a
different
resource,
type
and
I,
don't
know,
maybe
providing
a
reference
to
the
type
and
then,
but
that
would
the
downside
if
we
would
go
with
with
the
controller
reference,
you
would
have
to
read
the
actual
part
from
a
part-time
funding
from
the
reference
controller.
A
On
on
those
copy
and
then
eventually
provide
the
prospect
as
a
as
a
fallback
option,
if
you
cannot
read
that
information
off
of
the
cluster
anywhere.
B
So
we
thought
about
many
possible
use
cases,
including
using
pod
templates
that
are
also
like
a
top
level
objects
in
kubernetes
and
referencing
them
in
the
API,
and
that
creates
various
types
of
problems
like
the
sports.
Works
can
can
be
changed
and
they
need
to
be
monitored.
If
you
want
to
monitor
possible
objects,
like
I,
know
job
deployment,
then
you
expand
the
number
of
watches
that
are
required
to
implement
these
features.
B
So
this
all
of
these
things
that
you
mentioned
are
possible
are
not
rocket
science,
but
probably
it
would
be
better
to
do
the
after
MVP
and
after
initial
launch
and
see
with
this
like
real
or
strong
need
for
them,
because
they
will
complicate
the
picture
and
implementation
and
maybe
like
people
can
leave
with
copying
the
the
templates
and
in
that
way,
make
the
life
of
API
implementers
a
bit
easier.
B
But
anyway,
if
you
have
strong
feelings
about
the
that,
it
should
go
other
way
around,
please,
please
add
them
to
to
The
Proposal.
A
Daniel
one
thing:
the
presentation,
the
slides
that
you
shared
I
tried
and
I
cannot
access?
Can
you
share
that
with
the
batch
work
group,
mail.
C
B
C
Okay,
I
will
quickly
like
after
the
meeting
update
the
link
in
the
in
the
docs.
A
A
I,
don't
hear
anything
I'll,
probably
just
bump
this
topic
over
to
the
next
time
and
does
anyone
else
have
any
other
discussion
points
that
they
want
to
bring
up
with
the
group.
A
Okay
hearing,
none
with
that
I'm
gonna,
give
you
back
about
15
minutes
the
first
time.
Thank
you
very
much
for
today
and
see
you
next
time.
Bye.
All
right.