►
From YouTube: Kubernetes SIG Scheduling Meeting - 2019-12-19
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
Start
recording
hi
everyone
thanks
for
joining.
As
you
all
know,
this
is
a
public
meeting
ants
recorded
and
it
will
be
uploaded
to
the
to
YouTube.
So
we've
got
two
items
on
the
agenda,
so
the
first
one
is
Matthew
from
batch
on
GCE
he's
here
to
introduce
us
to
Google's
new
system
that
enables
patch
on
on
kubernetes.
A
B
Yeah
I'm
definitely
happy
to
wait
till
after
the
holidays.
Maybe
I
mean
especially
there's
any
discussion
or
questions,
then
that
makes
it
easier
rather
than
having
the
recording.
So
yeah
I
can
wait.
I
I'm
out
for
like
about
three
weeks,
so
it'll
be
kind
of
mid-january,
but
still
ok,
yeah.
How
do
we
do
that?
Yeah.
A
I
mean
we
still
got
a
lot
of
time
today
way
and
Mike
like
two
of
the
most
active
members
of
the
community,
and
they
have
strong
opinions
and
how
the
scheduler
should
go.
There
are
directions,
etc.
You
know
like
way,
is
from
IBM
Mike
from
Red
Hat,
so
it
might
be
a
good
idea
even
now
like
to
just
give
a
like
really
high
level.
What
is
what
is
K,
Bachchan
and
but
yeah?
Basically,
informal
discussion,
yeah
yeah,
sure
yeah
sounded
we
and
Mike.
B
Sure
so
I
so
I've,
just
a
few
slides,
so
maybe
I
can
start
with
that.
I
won't
go
through.
Maybe
all
the
details
this
time,
but
just
like,
maybe
the
first
few
just
to
kind
of
introduce
things.
Let
me
see
if
I
can
share
that.
B
B
B
One
kind
of
a
unique
feature
we
added
is
around
cost
management,
so
admins
can
set
things
like
budget
controls
and
we
also
support
auto-scaling
with
the
custom
auto
scalars.
So
we
can
minimize
the
amount
of
idle
resources
in
the
system
and
also
support
right
sizing.
So
we
tailor
the
size
of
the
VM
to
the
requirements
of
the
users
job.
B
B
C
B
That's
queues
are
generally
created
by
the
administrator
in
the
system,
so
you
could
imagine
having
say
a
batch
queue
for
each
team
or
different
batch
queue
for
different
types
of
priorities,
so
like
a
low
priority
queue
and
a
high
priority
queue,
and
that
leads
into
batch
priority.
So
on
each
batch
queue
you
can.
You
can
designate
the
priority
that
that
queue
will
have
and
what
we
do
is
we,
then
the
goal
is
then
to
attach
that
priority
to
every
key
task
that
runs
under
that
batch
queue.
B
And
finally,
we
have
to
support
our
our
budget
and
cost
management.
We
have
resources
which
are
batch
budget
and
batch
cost
model,
so
a
batch
budget
can
be
attached
to
a
queue,
and
so
everything
that
runs
into
that
queue
will
consume
from
that
budget
and
about
cost
model
is
used
to
define
how
much
each
resource
in
the
system
costs.
B
B
And
you
can
imagine
it's
it's
a
little
tricky
to
you
know
manage
budget
when
you
don't
really
know
in
advance
how
much
resources
a
job
and
I
use
and
so
kind
of
the
way
we
work
around.
That
is.
We
allow
the
users
to
specify
a
maximum
runtime
on
their
job
and
that
will
tell
us
the
maximum
cost
that
their
job
will
consume
and
that's
what
we
use
to
determine
if
a
job
will
be
allowed
to
run.
B
A
B
A
B
B
So
the
current
image
is
would
not
necessarily
so
we
definitely
don't
support
them
right
now,
outside
of
G,
ke
and
I.
Think
there
may
be
some
issues
simply
because
of
the
way
auto
scaling
works
in
our
system.
I
think
it's
kind
of
tied
fairly
closely
to
gke,
so
I
suspect,
running
it
outside
of
GK
right
now
would
not
work
very
well,
but
certainly
you
know
it's
certainly
on
our
radar
to
make
this
work
on
on
other
systems
and.
B
Yes,
so
we
have,
we
have
our
own
scheduler,
and
so
that's
we
still.
We
still
kind
of
go
through
the
standard,
culinary
scheduler,
but
we
set.
We
set
the
specific
node
name
on
each
pod,
and
so
we
kind
of
control
which
pods
are
sort
of
which
nodes
the
pods
get
assigned
to
yeah.
A
A
Scheduler
is
not
gonna
make
anything
like
it
was
just
this.
This
is
how
second
schedulers
work.
They
sit
the
node
name,
okay,
so
no
no,
the
schedule
logic
runs
on
it.
It's
not
even
considered
a
part
to
be
scheduled,
but
the
scheduler
default
schedule
basically
add
the
part
to
our
internal
cast.
Just
the
cash
just
to
you
know
account
for
the
resources
that
it
took
from
the
nodes,
but
we
don't
make
any
decisions
on
that
part.
It's
not
going
to
go
through
our
schedule,
exactly
that's
right.
A
A
E
D
B
C
B
D
I
think
yeah
next
time
you
feel
give
detailed
introduction,
I
think
use
some
some
audience,
whereas
the
next
question
was
a
difference
between
chemistry
and
never
changing.
Okay,
I
am
now
the
question
so
they're
like
okay.
No,
because
the
right
now
didn't
I
can
the
mechanism
there
you're
running
is
running
its
own
schedule.
D
Running
multiple
schedule
is
not
supporting
well
and
I.
Don't
think
in
the
in
the
in
the
near
term.
It
can
be
supported.
Well,
so
that
can't
let
never
be
a
problem
is
what's
the
real
map
of
K
badge?
Are
you
going
to
steal
and
going
in
this
direction?
I
mean
still
deployed
as
a
narrow
schedule,
or
you
do
have
some
plans
to
like
use
the
scanner
for
work
yeah.
We
factor
some
of
the
code
wrong
expanse
of
that
schedule.
Yeah.
B
C
B
Kind
of
separating
nodes
that
we're
kind
of
we're
kind
of
tainting
notes
that
we
want
to
use
for
our
systems,
so
they
can't
be
used
things
that
are
scheduled
or
scheduled
directly
through
the
regular
scheduler.
So
we
don't
end
up
with
those
types
of
conflicts
where
you
might
try
and
schedule.
You
know
two
things
on
the
same
note
and
they,
if
a
conflict.
B
A
B
Yeah,
that's
I
certainly
think
that's
possible
I.
Guess.
One
question
is
whether
you
know
we
want.
You
know
whether
batch
job
would
replace
job
necessarily
or
whether
would
it
would
exist
independent
as
additional
way.
You
can
run
a
job.
I
I
mean
so
one
issue
is:
if
someone
doesn't
care
of
a
batch
and
they
just
want
to
run
a
job
they
may
not
want
to
be.
You
know
having
to
deal
with
the
extra
the
extra
fields
that
are
kind
of
specific
to
batch.
A
C
A
C
E
C
D
Yeah,
just
just
one,
there
are
some
items
in
my
play:
twice
the
refactor
the
in
eliminate
the
failure,
reason
of
in
error
and
them
may
be
new
sometime
I
hope
today,
I
can
send
a
PR,
because
I
could
play
depends
on
some
of
our
codes.
So
I
want
to
be
cautious
on
just
in
fact
week.
It
is
more
and
another
is
I
need
to
sort
of
eliminate
the
register,
feed
predicate
or
something
those
kind
of
things
or
feed.
A
D
A
A
The
failure
reason
is
I
think
it's
definitely
a
priority
now
cleans
up
the
core,
the
generic
scheduler
and
allows
us
to
remove
the
predicates
and
priorities
from
inside
the
course
scheduler
right
and,
basically
and
and
the
predict
and
priority
metadata
producer,
I
create
a
bunch
of
issues
around
moving
the
the
code
of
the
various
predicates
and
priorities.
So
hopefully
this
will
move
in
in
parallel
and
and
finish
before
the
end
of
1.18
cycle,
and
then
we
can
celebrate.
We
can
focus
on
adding
more
features.
Hopefully
we
come
up
with
new
ideas.
A
How
what
plugins
said
tonight?
The
community
had
multiple
ideas
that
was
proposed
during
cone.
We
can't
pay
more
attention
to
that
and
improve
the
scheduler
functionality
after,
like
we
spend
I
think
now,
almost
five
six
months
on
basically
refactoring,
the
scheduler,
which
I
think
is
a
really
good
thing
and
it
was-
was
something
that
had
to
be
done.