►
From YouTube: Kubernetes WG Batch Weekly Meeting for 20221013
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
This
meeting
is
being
recorded.
All
right,
I'll
go
ahead
and
start.
B
Let
me
let
me
just
introduce
yeah.
Sorry
just
give
me
one
second
hi
everyone.
So
this
is
a
batch
working
group
meeting.
This
is
the
October
13th,
so
this
meeting
is
recorded
and
will
be
uploaded
to
YouTube
Just
disclaimer
and
please
adhere
to
kubernetes
code
of
conduct.
B
So
first
one
of
them
is
like
atna,
sorry,
Kevin
again
from
G
research,
I,
think
you'll
be
introducing
or
presenting
his
work
on
interactive
jobs
using
Armada,
but
before
that,
I
would
like
to
announce
or
remind
remind
everybody
of
kubecon
it's
coming
up
batch
day
on
October
24th,
which
is
not
next
week
the
week
after
that,
and
so
I
was
also
thinking
that
we
could
cancel
the
working
group
meeting
on
that
week,
because
most
of
us
will
be
in
kubecon
and
we'll
be
and
we're
meeting
there
and
I
hope.
B
Also
during
that
meeting
as
well,
you
can
see
like
magic,
Diana
and
other
people
from
the
community.
We
can
discuss
the
roadmap
for
the
back
working
group
if
we
have
like
a
sketch
of
the
document
by
that
time,
it
would
be
a
great,
like
you
know,
opportunity
to
discuss
it
face
to
face.
B
Okay,
so
with
that
out
of
the
way
I'm
gonna,
add
it
to
Kevin
Kevin.
Do
you
want
to
present.
A
A
Okay,
all
right
so
I
have
I,
have
a
Google
doc
that
I
added
to
the
the
agenda,
so
I
would
say
that
Armada
is
kind
of
an
inspiration
for
this
project.
A
In
soag
research
we
are
primarily
responsible
or
there's
a
lot
of
stuff
that
goes
on
there,
but
I
work
on
the
open
source
division
I'm,
mostly
on
Armada,
which
is
a
multi-cluster
kubernetes
scheduler
for
running
batch
jobs,
and
we
have
a
lot
of
users
that
like
to
do
different
things
and
one
of
one
of
the
things
I
found
they
they
tend
to
do.
Is
the
ability
for
running
Jupiter,
notebooks
onto
Armada
and
then
be
able
to.
A
You
know,
access
a
notebook
while
the
job
is
running
in
their
cluster
and
Armada.
Does
this
kind
of
pretty
they
have
a
pod
a
service
and
Ingress?
And
our
Armada
controller
basically
starts
all
these
together
and
then
the
user
can
use.
The
browser
to
you
know,
go
to
the
notebook
and
then
run
whatever
they
want
inside
their
python
code.
So
I
was
trying
to
take
a
look
at
what
Hugh
was
doing
the
batch
working
group
and
trying
to
do
this
using
a
job.
A
A
service
and
Ingress
and
then
I,
think
someone
suggests
the
idea
of
creating
a
custom
resource
definition,
and
so
I
have
a
just
a
little
proposal
here
that
I'll
briefly
walk
through
and
then
I
can
give
a
quick
demo.
This
will
be
a
very
short
talk,
but
please
feel
free
to
ask
questions
if
you
have
any
I
in
a
previous
job.
A
I
was
working
at
the
National
Institutes
of
Health,
and
one
of
the
one
of
the
popular
requests
was
actually
a
lot
of
times
with
researchers
using
Jupiter
notebooks
and
also
be
able
to
run
interactive
visualization
jobs
or
even
when
I
was
writing
some
parallel
code.
It
was
nice
to
be
able
to
get
a
cluster
to
run
a
debugging
tasks
when
you
need
a
chance
to
debug
something
live,
so
this
was
always
kind
of
a
common
use
case.
A
A
Okay,
yeah,
so
in
general,
there's
a
lot
of
different
use
cases
in
HPC.
It's
actually
I
got
I
didn't
list
a
few,
but
I
only
list
a
few,
but
at
I
know
at
NIH
there's
they
have
a
huge
section
about
actually
accessing
Jupiter
and
how
they
allow
users
to
do
that.
And
then
you
can
look
at
some
other
options
here
of
just
like
mostly
use
cases
where
people
do
allow
users
to
use
Jupiter
inside
of
their
research
clusters
and
yeah.
A
So
for
Armada
we,
basically
we
when
we
submitted
a
quote-unquote
job.
It
really
is
a
pod,
an
Ingress
in
a
service,
and
it
goes
to
one
of
our
kubernetes
worker
clusters
and
then
the
their
worker
cluster
is
responsible
for
starting
the
pods
starting
the
Ingress
in
the
service
and
tearing
those
down
once
the
once.
The
job
is
complete
and
then
use.
A
The
Ingress
is
obviously
used
for
act
for
exposing
a
URL
to
the
the
user
and
then
yeah
and
then
once
the
job
is
complete,
it's
removed
all
the
objects
are
removed.
So
really
why
I'm
here
is
I,
wanted
to
talk
a
little
I'll
give
a
demo
of
what
I
did
with
this.
A
The
the
custom
resource
definition
and
then
I
would
like
to
kind
of
maybe
talk
about
next
steps
for
integrating
with
Q,
if
possible,
because
I
I'd
like
to
be
able
to
see
how
this
would
work
with
a
natural
with
the
queue
as
a
queuing
solution,
yeah
and
so
I'll
just
go
to
the
demo
or
I'll
have
a
little
yeah,
whatever
it's
pretty
clear,
I'll
give
a
demo
now.
A
So,
basically,
this
is
my
first
time
creating
a
custom
resource
definition.
So
this
was
also
a
fun
little
project
for
me
just
to
understand
the
cube
Builder
and
how
to
generate
all
this
stuff
and
then
walk
through
some
of
this,
but
so
what
I
kind
of
so
my
my
specification
for
my
crd
is
a
a
job
template
currently
for
this
demo.
I,
don't
really
have
the
Ingress
one
working
I'm
using
kind
and
I
find
trying
to
debug
ingression
kind.
A
It's
never
my
favorite,
so
I
I'm,
just
gonna,
do
a
port
forwarding
for
now
and
so
I
have
a
service
and
I've
already
created
the
interactive
job.
A
So
you
can
see
I
have
this
Jupiter
and
it
it's
actually
running
a
job
that
is
still
running
and
then
the
Pod
is
there
and
then
I
have
a
service,
so
in
general,
I'll
just
support
forward.
D
A
So
it's
pretty
I,
don't
have
it
very
clean
or
nice
right
now.
So
it
is
mostly
going
to
show
you
the
login
for
for
Jupiter
and
then
there's
a
token
that
you
can
read
from
the
logs
to
get
it
working
but
I,
just
kind
of
wanted
to
show
that
now,
like
since
I'm
port,
forwarding
the
service
and
this
job
is
running,
I'm
able
to
I,
have
this
notebook
up
and
running
and
then
yeah
basically.
A
What
was
I
was
gonna
say
so
the
the
spec
is
just
I
have
a
job
template
which
allows
you
to
run
the
job
and
then
for
now,
I
just
have
a
I
hard
coded
to
be
a
node
Port
service
and
it
listens
on
the
same
port
as
the
container
and
then
part
of
the
the
controller
responsibility
is.
It
starts
a
job
and
a
service
and
then
once
I
remove
the
the
Jupiter
the
interactive
job.
It
removes
other
resources,
so
I
can
just.
A
So
then,
you
can
see
here,
service
yep
everything
was
deleted,
so
I
guess
this
is
I
mean
it's
a
pretty
short
demo,
but
so,
like
really,
my
main
questions
for
this
group
are
about
like
I
guess
what
I
wanted
to
show.
This
is
because
I
I
found
that,
with
working
with
research
groups,
a
lot
of
times
batch
jobs
aren't
a
lot
of
times.
A
People
do
want
some
kind
of
networking,
especially
for
like
Jupiter
and
I
notice,
that
a
lot
of
Q
is
focused
on
the
job,
controller
and
I
wanted
to
know
about
if
it's
possible,
for
integrating
something
like
this
into
queue
and
how
we
could
go
about
that
and
that's
kind
of
the
only
the
question
I
have
really
about
it,
for
this
is
just
talking
about
like
what
the
steps
would
be
for
that
and
any
questions
you
have.
B
I
have
two
questions,
but
I'll
leave
it
first
to
the
Community
if
there
are
other.
If
there
are
questions.
C
Yeah,
so
listen
thank
you
Kevin
for
sharing
this.
This
is
really
interesting,
so
it
sounds
like
it's
part
of
the
requirements.
If
you
want
to
be
able
to
express
holistically
in
some
aspect,
multiple
resources,
not
just
the
job
right,
so
that
the
whole
all
the
set
of
resources,
can
be
dispatched
or
removed
from
the
system
holistically.
Exactly
is
that
kind
of
the
basic
requirement
for
this
workload.
C
And
then
also
since
these
don't
really
run
to
completion
which
I
think
of
as
jobs
and
the
general
sense
it's
it's
really
a
standing
up
of
a
pot
and
then
there's
explicit
removal
of
these
objects
in
the
sense
of
when
you
do
delete.
You
want
to
delete
all
the
objects,
but
they
don't
ever
run
to
completion.
Is
that
correct.
A
Yeah
my
my
thought,
for
that
is
a
Timeout
on
the
the
queuing
either
as
part
of
the
queue
or
as
part
of
the
the
job,
but
then
I
guess
I
need
to
see
if
the
timeout
would
also
Delete
the
the
crd
completely
so
I
don't
know.
I
haven't
really
thought
about
that,
but
that's
at
least
kind
of
what
we
thought
about
is
at
least
the
Pod
will
be.
It
has
some
kind
of
active
deadline
that
will
terminate
after
a
certain
amount.
D
Yeah
so
I
I
have
a
similar
question,
but
I
have
another
one
as
well,
which
is
you
demoed
with
the
notebook
but
then
actually,
because
you
were
talking
about
Ingress,
but
actually
you
just
did
put
forward
so
in
in
principle.
An
interactive
job
should
be
just
like
layer.
Four,
it
doesn't
have
to
be
like
HTTP
forward
living
right,
like
the
Jupiter
example
implies
Ingress,
but
for
the
generic
case
we
probably
are
not
talking
about.
Ingress
is
more
more
something
else
like
exposing
on
layer
for
the
port
directly,
not
at
the
hdb
layer.
A
Yeah
I
mean
in
in
Armada,
we
do
have
the
Ingress
stuff
working
I,
just
I'm
using
kind
and
I
always
find
it
troubling
to
get
my
Ingress
stuff
working
from
like
from
kubernetes
local.
So
it.
D
Could
be
like
a
generic
joke
where
you
just
either
get
a
shell
directly
on
the
remote
pod,
somehow
using
a
port,
it
doesn't
have
to
go
through
through
Ingress,
so
I
was
just
thinking
also
looking
forward
to
something
like
the
Gateway
API
with
where
you
have
both
layer
four
and
there
are
seven
possible.
Okay,
it's
probably
yeah.
A
D
A
Yeah
I
know
that
there
will
be
another
talk
about
that.
I,
don't
know
much
about
the
Gateway
API,
but
I'll
take
a
look
at
that
thanks
for
that,
but
yeah
I
think
in
general
I
would
want
it
to
be
in,
like
the
final
solution
would
be
working
with
an
Ingress
I.
Just
don't
know
exactly
what
kind
of
configuration
like
I
would
want
for
an
Ingress
at
this
point
so
for
the
demo
I
just
kind
of
skipped
to
that.
A
In
our
motto
we
kind
of
have
a
like:
we
use
the
name
of
the
port
and
then
that's
the
host
or
sorry,
the
name
of
the
Pod
that
creates
the
host
and
then
we
just
have
whatever
the
ingresses
and
Armada.
That
is
what
allows
the
users
internally
to
view
the
job,
that's
what
is
done
with
Armada,
but
that
would
be
the
final
solution.
Obviously,.
B
So
the
the
idea
of
queuing
and
it's
mostly
centered
around
Resource,
Management,
right
and
and
we're
talking
about
hard
resources
like
things
that
are,
you
know
you
have
limited
capacity
of
like
CPU
memory,
Etc
yeah,
no,
exposing
a
service
is
not
necessarily
A
Hard
resource,
unless
we
think
we
have
a
limited
number
of
ips
that
we
want
to
manage,
and
so
in
my
mind
this
is
not
necessarily
something
that
is
going
to
be
the
concern
of
the
queuing
operator
like
whether
or
not
we
create
a
service
and
expose
it.
B
So,
in
a
sense,
your
integration
with
Q
is
kind
of
orthogonal
to
the
requirement
that
this
job
when
it
starts
it
needs
to
be
fronted
by
a
service,
so
that
because
you
can
always
create
the
service
from
the
beginning,
it
doesn't
matter
right.
It's
not
taking
up
any
resources,
we're
not
trying
to
manage
anything
related
to
it,
but
once
the
job
actually
starts
it
will
become
responsive.
So
that's
the
only
side
effect,
but
creating
it
beforehand
after
the
drop
starts
Etc
it.
B
It's
not
really
the
concern
of
queuing
or
dynamic
quota
management,
and
so
it
is
going
to
be
really
the
concern
of
whatever,
like
you
know,
for
example,
in
your
case
Jupiter
like
you,
just
call
it
like
job
operator
that
creates
the
service
once
the
job
starts,
or
maybe
even
creates
it
even
before,
and
you
have
some
sort
of
I
don't
know
already
paying
or
whatever
like
that
kid.
Okay,
once
the
service
becomes
responsible,
it
means
that
the
job
was
scheduled,
and
now
it's
serving
you
know
the
the
request.
B
So
that's
one
thing
that
I
think
I
would
like
to
clarify
the
other
thing
related
to
Finnish,
so
I
guess
in
a
sense,
who's
going
to
finish
the
job.
It's
the
developer
or
some
other
right.
It's
not
like,
say:
okay,
I
looked
out
of
Jupiter
of
this
session.
That
means
the
job
job
has
ended,
and
then
one
during
interpreter
gives
that,
like
you,
know
ability
basically
okay
I
logged
off
and
so
that
job
should
tell
me
yeah.
A
I
think
that
is
something
that
I
I've
been
thinking
about,
and
I
I
posted
an
issue
on
cue
about
having
kind
of
I
guess
some
unclear
fully
on
the
the
nomenclature.
But
I
was
wondering
for
like
I.
A
Oh
sorry,
I'm
not
as
familiar
with
how
this
is
handled
in
like
an
HPC
Community,
but
I
have
seen
typically
when
people
have
interactive
jobs,
they
they
have
time
limits
on
them
and
I,
don't
really
I
guess
these
can
be
handled
by
a
user,
creating
like
the
job
as
an
active
deadline,
but
I
didn't
know
if
we
should
have
like
timeouts
as
part
of
a
queue.
A
So
that,
like,
like
you,
know,
I
imagine
there'll,
be
like
an
interactive,
cue
or
interactive
job
queue
which,
with
has
like
a
hard
deadline
of
you,
know
30
hours
and
then
that
at
least
forces
the
jobs
to
end,
because
I
do
think
yeah.
This
could
be
a
a
permanent.
If
that
you
don't
have
any
kind
of
time
limit,
you
can
have
somebody,
you
know
using
a
GPU
for
infinite
amount
of
time
for
whatever
they
want
to
do,
which
is
not
a
great
case
for
sharing
your
resources.
Among
a
group
yeah.
B
I
mean
that's,
that's
like
an
interesting
requirements
basically
like,
and
we
would
like
we're
discussing
preemption
and
until,
although
we'll
we'll
talk
about
this
in
a
minute,
this
can
be
sort
of
like
as
one
form
of
preemption,
but
not
there's,
no
new
job.
That's
coming
to
preempt,
it's
just
basically
you're
getting
preempted
because
of
a
deadline.
So
that's
an
interesting
like
you
know,
you
know
requirement
it's
not
really
related
to
interactive
jobs.
Right
like
it's.
A
genetic
requirement
around
okay
setting
up
deadlines
around.
A
D
B
Thing
that
you
might
want
to
look
at
is
like
with
job
you
can
have.
B
You
know
you
can
create
index
job
where
you
can
also
associate
it
with
a
headless
service,
and
that
would
expose
you
know
like
a
predeterministic
IP
address
like
a
record
for
for
each
index,
and
so,
if
you
create
a
one
part
job
and
create
a
headless
service
for
it,
you're
gonna
get
like
a
a
stable
DNS
name
to
that
part
which
is
going
to
be
the
job
name,
Dash
zero,
which
is
like
the
index.
B
So
this
could
be
a
pattern
that
you
might
want
to
apply.
Okay,.
B
But
it's
it's
not
necessary,
like
you
can
always
yeah.
If
you
have
your
own
CRT,
then
at
the
end
of
the
day,
you
can
do
whatever
you
want
right
like,
but
this
could
be
another
way
of
doing
this
without
necessarily
having
a
crd.
Basically,
every
time
you
create
the
job,
you
just
create
a
headless
service
for
that
job
and
that's
it
okay,
but
you
need
to
enforce
that
these
jobs
are
created
as
in
these
jobs.
A
B
E
A
Yeah
and
then
I
guess
Ricardo
Deer
previous
point
about
the
logging
into
a
node
I
I
was
I
was
thinking
about
that
one
for
a
little
bit,
but
I
I
figured
it's
a
little
at
least
it's
kind
of
weird
to
be
able
to
SSH
into
a
container
I
think
you
have
to
actually
add
like
SSH
capability
to
the
container
and
I
didn't
want
to
require
that
you
have
to
modify
containers
in
order
to
do
that,
but
I
guess
that
is
possible.
F
I'm
also,
maybe
you
want
to
look
at
well
if
I
just
forgot
the
name
now,
the
Ingress
is
replaced
by
another
API,
not
replace,
but
there
is
an
alternative.
Apiary
Gateway,
that's
the
word.
Yes,
okay,
you
might
want
to
look
into
that,
but
back
to
queuing,
I
wanted
I
wanted
to
mention
that
since
you're
using
the
job
API,
if
you
just
install
install
queue
on
your
cluster,
those
those
embedded
jobs
will
already
be
queued.
F
If
that's
not
the
desire,
then
then
that's
where
we
need
extra
integration.
So
basically
you
want
to
have
your
crd
also
have
a
suspend
feel
so
whenever
Q
decide,
that
is
ready.
You
can
flip
that
flag
so
that
that's
kind
of
the
mechanism.
A
You're
saying
that
I
mean
so
right
now,
like
sir,
like
you
could
already
create
the
service
and
Ingress
and
then
add
your
workload.
The
job
suspended
true
and
all
that
would
probably
work,
as
is
the
only
issue,
is
cleaning
up
the
job
and
service
after
the
job
is
complete
right,
which
is
a
manual
process.
F
Okay,
so
it's
it's
more
about
whether
you
care
about
the
cell
is
being
created
from
the
beginning,
so
that
that
would
be
the
difference,
although
you
could
make
it
you
could
make
it
so
that
the
crd
doesn't
create
the
service
until
the
job
is
marked
as
suspend
equals
false,
which
could
be
your
your
your
gain
weight
for
you
control
your
control
flag
for
for
the
other
two
objects.
So
that's
a
possibility
too.
F
That
would
be
a
fairly
easy
integration,
but
so
that's
all
good
and
that
all
works
when
your
job
is
already,
you
know
a
fixed
size
like
you
know,
you
know
you,
you
just
want
one
part
throughout
the
entirety
of
the
the
the
interaction.
F
But
the
question
is
what
happens
when
if,
if
the
in
Notebook,
you
can
create
extra
workers,
sources,
I,
don't
know
if
that's
possible,
I'm
guessing
it
is
possible.
Yeah.
A
I
mean
I,
guess
like
the
idea
would
be
like
for
when
I
think
about
like
with
the
normal
HPC,
like
you
could
say:
I
want,
you
know
four
nodes
and
then
or
sorry
you
know
four
dedicated
computers
in
order
to
run
like
an
MPI
job
for
debugging
purposes
within
those
four
nodes.
So
I
think
that,
like
that
would
be
a
requirement
down
the
line
for
this
is
that
you
would
probably
want
to
be
able
to
use
more
than
I
guess:
I
get
a
little
confused
with
pod
and
nodes.
D
F
Multiple
MPI
notes:
yes,
usually
those
would
have
to
be
Representatives,
multiple
Bots,
but
that's
not
really
a
problem
for
Q.
If,
if
you,
if,
if
the
number
of
workers
is
fixed
from
the
beginning,
that's
not
a
problem.
What
is
missing
is
the
more
like
a
dynamic
scenario
where
you
start
with
one
part,
and
then
you
require
at
some
point
different
number
of
parts,
and
then
you
scale
down
and
scale
up
and
so
forth.
That
kind
of
interaction
is
not
defined.
F
There
is
some
people
trying
to
to
to
build
this
into
queue
button,
but
there
are
some
some
things
missing
so
that
that's
where
more
work
might
be
needed,
but
okay
size,
even
if
multiple
pots-
that's
that's
already
supported.
A
Okay,
I
will
play
around
more
with
fan
if
I
can
just
integrate
this
with
Q
directly
and
then
I'll
think
about
more
about
the
service
and
Ingress
like
the
life
cycle
for
that,
but
yeah
thanks
for
everyone
for
their
time.
I
know,
although
you.
C
Okay,
one
more
question:
it
was
really
more
for
the
community,
and
that
is
thinking
forward.
This
I
think
the
concept
of
identifying
not
just
the
Pod,
creating
resources,
but
other
resources
and
kubernetes
holistically
is
a
powerful
concept
and
I'm
wondering
down
the
road
being
able
to
take
advantage
of
that
Concept
in
queue.
Do
we
see
this
is,
is
something
we
might
want
to
expand
on
this
kind
of
requirement.
C
C
Is
that
something
we
might
think
through
or
as
powerful
enough,
and
this
is
an
example
of
a
use
case
where
it
would
be
I'm
thinking
individual
would
be
nice
to
have
here's,
here's
all
the
objects
that
I
need
to
represent
my
job
or
my
my
workload
and
cue
it,
and
then,
once
the
you
know,
once
it's
ready
to
get
dispatched
based
on
you
know,
whatever
the
requirements
are
resource
demand
then
dispatch
all
the
workloads
within
that.
C
C
C
I
was
gonna,
say
I'm,
just
thinking
of
it
as
right
now,
Q
is,
is
queuing
a
the
job
resource
and
I'm
wondering
as
long
term
as
we
think
about
cueing
and
taking
this
concept
of
holistically,
representing
not
just
the
the
one
job
creating
prod
object,
but
multiple
resources.
If
we
think
that's
something
that
we
can
that's
a
powerful
enough
to
start
to,
you
know
think
through
what
we
might
do
as
far
as
cueing.
All
these
objects
that
represent
a
job.
B
Right
so
cute
doesn't
really
cue.
Jobs
accused
something
called
workload
and
it's
a
abstract
concept
right.
That
represents
a
collection
of
resources
that
you
have
limited
capacity
of,
and
you
want
to
subject
them
to
quota,
and
so
whether
or
not
those
represent
a
job,
two
jobs,
a
job
and
a
deployment
is
really
up
to
the.
Let's
call
it
the
workload
controller
right.
B
B
It's
not
that
it's
just
tells
Q
how
much
resources
I
want
and
then,
once
these
resources
become
available,
tell
me
to
start
and
when,
when,
when
you
tell
that
workload
controller
to
start,
it
will
create
them
as
whatever
it
wants,
and
so
in
that
sense,
what
you
mentioned
is
kind
of
already
supported
it's
again
up
to
the
batch
workload,
controller.
B
B
B
I
mentioned
at
the
beginning:
what
is
what
is
it
that
you
want
to
control
with
the
service
is?
Is
there
a
resource
that
is
associated
with
the
service
that
you
want
to
basically
control
and
subject
to
Resource
quota?
That
isn't
really,
but
I
mean
it
could
be
that
oh
I
don't
want
to
create
the
service
until
my
job
is
scheduled
right.
So
this
is
not
really
you're,
not
really
associating
it
with
a
with
a
quota.
C
A
Mean
I
guess,
but
when
ports
and
services
be
considered
service
like
constrained,
if
you
have
like
for
like
Armada,
you
know,
we
have
a
lot
of
users
and
we
can't
have
everybody
open
up
a
port
and
not
clean
up
that
port
and
So.
Eventually,
you
don't,
you
run
out
of
ports
to
use.
Also,
if
you
just
let
people
schedule
them
whenever.
B
So
so
in
kubernetes
I
mean
like
IP
management
is
happens
in
a
much
different
layer
right
and
it's.
It
is
not
like
a
job
level
resource.
Really
it's
it's
a
it's
related
to
the
pod.
B
So
that's
one
thing,
but
if
yeah
I
mean
I,
I
port
forwarding
is
again
like
again:
it's
yeah
I,
don't
think
it's
an
encouraged
like
it's.
It's
a
pattern
that
in
general
the
community
encourages
like
to
do
port
forwarding
on
the
Node
I
mean
it
has
security
implications
and
whatnot.
It
needs
to
be
like
you
know,
you
need
to
set
up
like
a
proper
Gateway
forwards,
things
to
like
back-end
part,
Etc,
I,
don't
know
if
that's
something
you
want
to
manage
as
a
resource.
It's
too
complicated.
A
I
guess
yeah.
My
point
is
that
eventually,
for
a
large
enough
research,
cluster
I
would
say
ports
do
or
are
something
you
want
to
have
some
control
over,
because
you
don't
like
you
do
want
to
make
sure
that
they
get
once
that
job
is
complete.
I
would
say
you
probably
want
those
ports
open
again.
So
that's
always.
B
D
B
But
again
like,
if
you
set
them
up
properly
using
like
you
know,
Ingress
and
whatnot
up,
like
you
know,
it's
really,
like
a
part,
is
associated
with
the
IPO
virtual
IP
that
you
assigned
to
that
service.
It's
not
a
limited
resource.
It
goes
themselves
all
of
themselves
like
it
depends
on
the
system.
A
D
B
D
F
Perfect,
thank
you
so
yeah.
D
F
Me
Maybe,
if
I
try
the
entire
screen,
it
works.
Oh,
let
me
try
that.
D
F
Thank
you,
okay,
sharing
that
doesn't
work
so,
okay
in
queue.
We
are
we're
working
on
a
proposal
for
Warcraft
preemption.
This,
basically
two
scenarios
where,
where
preemption
might
be
needed
in
Q
There,
is
a
concept
of
borrowing
quota.
F
So
if
you
set
up
quota
for
a
cluster
queue
and
then
multiple
cluster
cues
are
set
up
to
to
be
in
the
same
cohorts
in
the
same
group,
they
can
share
quota
when
rather
one
cluster,
you
can
borrow
quota
from
another
if
this
other
cluster
queue
is
not
using
the
quota
now,
the
problem
is
how
to
recover
the
quota.
Currently,
the
only
way
is
to
wait
for
the
workloads
to
finish,
and
this
is
clearly
not
enough
for
certain
use
cases.
F
We
want
to
basically
end
workloads
faster
well,
the
other
use
cases
within
a
cluster
queue.
You
could
have
a
a
high
priority
job
coming
in
and
even
though
you
you're
not
borrowing
quota
from
other
cluster
queue,
you
still
might
want
to
preempt
a
a
workload
with
lower
priority.
So
that's
that's
the
use
case.
I
I
think
it
should
be
relatively
common
concept
across
patch
schedulers,
so
well.
F
The
proposal
I
was
kind
of
being
I
was
trying
to
be
as
comprehensive
as
possible
or
give
users
the
a
good
amount
of
knobs
to
to
influence
preemption,
so
one
kind
of
Knob
I
was
trying
to
give-
and
this
is
based
on
on
some
discussion
that
was
happening
in
the
issue.
If
you
go
to
issue
83
4q,
there
was
some
some
requests
to
be
able
to
favor
the
preemption
of
certain
workloads
based
on
how
long
they
have
been
running
now.
F
I
came
up
with
a
kind
of
maybe
a
little
bit
too
complex
API.
That
kind
of
sets
cost
a
cost
curve
for
the
time
based
on
the
time
or
how
how
good
or
how
bad
it
is
to
preempt
a
workload
given
how
long
it's
been
running.
So
this
is
my
proposal.
I
think
the
most
of
the
feedback
feedback
I've
gotten
so
far-
is
that
it's
maybe
too
much
to
start.
Maybe
we
can
just
get
rid
of
it
until
there
is
more
justification
for
it.
F
So
if
you
think
this
is
something
you
might
be
able
to
use,
let
us
know
otherwise
we
will
park
it
for
now
and
the
other
one
is
maybe
a
little
bit
less
morning
intuitive.
Basically,
when
your
workload
wants
to
pre
needs
to
preempt
some
other
workloads,
just
simply
saying
how
long
it's
okay
for
you
to
wait
for
other
workloads
to
finish
before
you
start
pre-enging.
F
So
this
could
be
useful.
If
you,
you
know,
you
have
short
running
jobs
and
then
you're,
okay,
waiting
for
the
period
this
period
of
time
and
so
yeah.
Basically,
that's
the
API
I
I
have
another
proposal
over
here,
also
to
give
extra
control.
F
What
is
it
like
workload
sorting
so
how
how
to
sort
the
workloads
they
are
already
running
to
determine
which
ones
are
the
best
ones
to
to
preempt?
Honestly,
I
think
this
is
probably
Overkill
at
the
moment
so,
but
I
wanted
to
leave
it
here
so
that
you
more
or
less
understand.
What's
the
the
current
proposal
for
how
workloads
should
be
chosen
like
first
sort
by
lowest
priority,
then
maybe
a
short
lowest
cost
as
proposed
earlier
and
then
by
lowest
running
time.
F
Yeah
I'll
just
finish
up
real
quick.
F
The
API
proposal,
I'm
still
working
on
the
actual
algorithm
I,
hope
I
hope
to
to
have
some
this
completed
for
kubecon.
So
if,
if
you
have
a
chance
to
talk
with
me,
you
we
can
discuss
it
or
you
can
of
course
submit
feedback
in
this
in
this
PR
already
yeah.
That's
it
yes,
Diana.
C
Oh
sorry,
I
didn't
mean
to
interrupt
you.
There
I
just
had
a
quick
question.
So
I
haven't
read
through
this,
but
do
you
have
the
basic
concept
of
expressing
whether
a
workload
or
is
preemptable
or
not,
meaning
that
if
there's
borrowing
you
know-
and
you
want
your
job
to
run
you
and
it's
not
a
preemptable
or
you
don't
want
it
preempted,
then
that
allocation
of
the
quota
would
not
be
part
of
the
preemptable
type
of
resources
that
would
they
can
get
reclaimed.
C
So,
in
other
words,
you
know
I
say:
okay,
I
want
to
borrow
from
my
workload,
I,
don't
mind
if
I,
if
I
need
to
borrow
I,
borrow
and
I
know
that
I
have
the
risk
of
getting
pamped
in
right
and
I
can
set
my
parameters
for
that.
Okay.
F
That's
a
good
point:
there
is
no
switch,
not
what
the
most
you
can
do
is
the
the
knob
is
for
the
cluster
queue
rather
yeah.
Exactly
if,
if
the
cluster
queue
is
like,
it's
used
usually
has
workloads
with
this
kind
of
requirements,
then
you
can
simply
disable
borrowing.
Look
for
it
and
I
think
this
kind
of
things
is
better
to
leave
to
the
administrator.
F
So
setting
up
setting
it
up
in
the
cluster
queue
might
be
a
better
fit
because,
if
you
put
in
the
workload
it
might
lead
to
a
abuse
right
of.
C
Course,
everyone
knows
that's
true,
but
it's
true,
but
I
also
think
it's
a
powerful
as
well
that
that
I
I
know
if
I
submit
this
job
and
it
gets
accepted
and
with
meeting
all
the
policies
that
I
know
it's
not
going
to
be
preempted
but
I
have
this
other
job
that
I
don't
mind
running
in
the
cluster.
C
You
know:
I'm
the
same
user
I'm
using
the
same
quota,
but
I
have
this
other
job,
that
I
don't
mind
it
getting
preempted
and
so
being
able
to
express
that
within
the
one
job
adds
more
flexibility.
So.
F
I
just
wanted
to
know
if
you're
I
think
you
can
achieve
exactly
that
by
setting
to
Cluster
cues,
where
the
first
cluster
queue
has
the
real
quota
and
the
second
cluster
queue
has
zero
quota,
and
it
can
only
borrow
from
the
front
existing
classic
queue
and
other
cluster
Keys,
so
that
could
be
a
way
to
set
that
up,
giving
all
the
control
to
the
administrator
instead
of
of
the
user.
Of
course,
the
user
has
a
chance
to
use
one
classic
or
the
other.
F
C
F
Yes,
I
think
I
had
a
better
thought.
While
you
were
talking
about
this,
but
I
forgot.
C
F
F
F
E
Yeah
sure
I
mean
I
was
just
you
know,
taking
a
point
on
on.
What
we
just
discussed
is
that
there
are
two
cues
one
can
borrow,
and
one
cannot
borrow,
but
you
have
like
n
jobs
skewed
in
a
in
a
cluster
queue
that
that
that
cannot
borrow
like
I
mean,
meaning
all
all
the
jobs
should
should
complete.
First
so
like.
How
do
you
prioritize
it?
In
that
scenario,
foreign.
F
F
The
workload
can
have
priority,
but
otherwise
it's
mostly
fifo
first
thing.
First
out:
okay,
okay
got
it.
Yes,
there
is
somebody
working
I,
don't
know
if
Kant
is
here
or
Alex
I.
F
Don't
remember
sorry,
I,
don't
remember
which
one
of
you
were
working
on
a
proposal
for
giving
more
knobs
for
sorting
winning
a
cluster
queue
so
yeah,
because
today
we
only
have
five
four
kind
of
a
strict
fifo,
which
is
five
all
the
time
or
some
fifo
where,
if
it
doesn't
fit,
it
still
tries
to
schedule
other
small
other
workloads.
F
But
there
was
an
idea
of
expanding
these
knobs
to
more
sorting
mechanisms,
but
that's
still
ongoing.
Okay.
B
All
right
five
minutes
over
time.
Thank
you,
everybody
and
hope
to
see
most
of
you
in
kubecon
in
two
weeks,
just
as
a
reminder,
I
think
we'll
be
canceling.
The
working
group
meeting
on
that
week
I'll
send
an
email
and
post
a
message
on
this
live
Channel
as
well.