►
From YouTube: Kubernetes SIG Scheduling Meeting 2018-05-31
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,
okay,
so
I
started,
recording
for
a
week,
I
actually
have
a
like
several
items
on
the
agenda
that
I
would
like
to
follow
up
one.
The
first
one
is
actually
sort
of
like
the
highest
by
your
idea
at
this
point
given
given
the
status
and
that's
actually
something
that
Harry
was
working
on
with
respect
to
image,
locality,
priority
function.
Harry,
do
you
have
any
update
on
that?
I
know
that
it's
kind
of
blocked
by
the
fact
that
images
are
not
reported
in
a
consistent
way
from
the
node
to
the
scheduler.
B
A
C
Yeah
I
am
mystic
note
about
this
issue
today,
and
we
have
reached
agreement
that
we
can
just
work
around
it
right
now.
I
ignore
some
of
match.
Example:
yeah
problem
go
darker
report,
a
shorter
image
name.
We
just
just
annoyed
because
you
know
it's
risk
to
receive
that
all
they
come
there
on
time.
They'll
report
consistent
engineer,
so
I
think
it's
so.
C
It's
just
like:
well,
we
will
only
say
to
the
image
name:
we
used
to
pull
the
image,
no
matter
what
can
go
wrong?
Have
we
use
faisal,
oh
yeah,
so
if,
if
some
container
100
port
something
different
from
its
from
the
negativities
used
to
pull
there,
we
just
ignore.
So
it
says
it
will
be
only
a
small
part
of
the
case.
Okay,
okay,
yeah.
Definitely.
C
A
If
there
is
anything
under
scheduling
side
that
you
would
like
us
to
review,
please
pay
me
or
maybe
Klaus
or
Jonathan.
One
of
us
can
do
the
review
basically
before
before
the
code
freeze,
I,
usually
you
know,
I
tried
for
a
while
to
keep
up,
be
the
notifications
that
I
received
from
github
about
the
code
reviews
that
are
assigned
to
me,
but
after
a
while
they
become
too
many
and
I
cannot
really
keep
up
with
the
notifications.
So
at
this
point
it's
just
basically
me
randomly
browsing
github
pages.
A
C
D
A
E
So
for
the
tenant
know
the
back
foundation
I
think
maybe
we
can
with
better
to
move
master.
It
is
that
because
they're
about
three
or
four
p
are
under
the
reveal
has
to
depend
on
the
feedback
from
the
reviewer
okay,
so
you
think
you're
not
gonna
make
it
to
111.
Yes,
I
think
that's
will
be
safely
to
four
nights
already
said
you
know
we
need
to
see.
You
know
this
review.
Jana
may
also
need
to
some
Kibo
Giulio
different,
a
component.
That's
Frank,.
A
A
E
B
A
And
there
is
actually
with
respect
to
priority
and
preemption.
There
is
another
PR
that
Ravi
sent
Rob
since
Rob
is
online.
We
can
ask
him
so
Ravi
I'm
talking
about
the
PR
that
enables
priority
and
preemption
by
default
and
empowerment
and
changes
are
end-to-end
tests
so
that
it
works
with
priority
and
preemption.
What
is
the
status
there?
I
know
that
it's
kind
of
blocked
by
by
the
other
PR
that
adds
priority
to
coda.
That
is,
you
have
any
update
from
the
Red
Hat
side,
yeah.
F
So
that
PR,
which
is
actually
blocking,
is
moving
forward.
So
there
is
one
top
level
approval
like
Clayton,
who
has
given
couple
of
comments,
and
we
have
addressed
his
and
Derek,
is
also
interested
in
moving
this
forward,
and
it
has
already
spoken
to
Clayton,
so
it
should
be
merged
before
the
end
of
this
week
or
before.
Fifth,
before
the
code
freeze,
okay,.
C
A
F
There's
one
more
PR
that
I
have
created
that
actually
got
blocked
by
this
fear.
So
I
spoke
to
David
each
from
Red
Hat
and
he
is
fine
with
the
change
that
we
are
making
so
I
just
wanted.
What
is
our
PR?
That's
the
that's
the
P,
and
that
actually
knows
to
beta.
So
we
have
two
beers
one
is
blocking,
which
is
because
it's
coated
priority
classes
to
quota
to
priority
classes.
F
A
The
only
thing
right
now
is
actually
the
the
the
fact
that
we
are
not
going
to
schedule
demon
set
parts
by
default
scheduler
in
111.
There
is
a
little
bit
of
a
concern
which
actually
I
had
initially,
because
I
was
thinking
that
this
is
a
prerequisite
to
moving
priority
and
preemption
to
two
beta,
but
now
I
feel.
Maybe
it's
not.
The
reason
that
I
was
thinking
that
it
is
initially
is
that
we
are
gonna.
A
Disable
who
is
scheduler
in
111
disabling
reschedule
means
that
if
there
are
critical
parts,
a
scheduled
reschedule
ur
is
not
going
to
create
room
for
that
and
I
felt
like
since
the
immense
a
controller
doesn't
do
for
a
preemption.
This
could
be
a
problem.
All
the
positive
is
scheduled
by
scheduler
or
fine,
but
parts
that
are
not
scheduled
by
scheduler
could
be
in
trouble.
However,
when
I
talk
more
about
it,
I
realized
that
demon
said
controller
would
not
create
cards
if
nodes
cannot
run
them.
A
So,
even
when
scary
scheduler
existed
in
the
past,
because
reschedule
was
not
helping
with
demons
at
pots,
getting
scheduled
or
critical
demons
at
Pascal
Scot
so
looks
to
me
that
reschedule
er.
The
fact
that
we
schedule
does
not
does
not
exist
in
111
shouldn't
be
making
any
difference
for
scheduling
demons
at
pods
anyway,
you're.
Basically
speaking
the
same
as.
A
So
but
you
know,
the
thing
is
that
regardless,
the
problem
is
the
ultimate
controller,
never
creates
immense
at
thoughts
until
it
realizes
where
it
finds
out
that
a
node
can
run
those
balls,
so
unreal
scheduler.
The
way
that
the
scheduler
works
is
is
that
it
actually
looks
at
the
pending
parts
of
the
critical
parts
and
then
goes
and
removes
some
some
parts
from
nodes
to
make
room
for
those
parts.
This.
E
A
B
A
E
E
A
Oh
yes,
yes,
you
should
probably
merge
this
PR
and
so
basically
I
left
two
comments.
The
first
one
was
that
since
you're,
removing
reschedule
this
is
gonna,
be
a
problem
if
we
don't
enable
it
by-
and
the
second
comment
was,
the
fight
was
what
I
just
basically
said
in
this
meeting-
that
maybe
reschedule
didn't
help
with
scheduling
Beaman,
surprise,
but
but
you're
right
actually
now
I
think
when
I
think
about
it.
I
know
that
the
simulaid
function
of
demon
said
demon,
dimension,
controller
and
so
for
four
critical
posits.
E
Feature
I
think
that
should
be
fine,
because
the
demon
this
feature
is
the
Alpha
and
in
the
alpha
test
in
the
Alpha
end-to-end
it
has
so
we
also
need
about
the
the
preemption
right
mm-hmm.
Yes,
so
I
think
a
way.
What
we
need
to
do
maybe
win
is
to
to
update
some
Yama
failed
to
make
the
grade
book
critical
polled
to
have
to
hire
her
party
to
how
her
party,
so
we
can
leverage
the
permission
to
or
for
the
for,
the
Antoinette
ask
how
the
Union
our
fact
has
the
environmental.
All
the
r32
are
enabled.
A
Now
but
the
problem
is
that
we've
not
talking
about
alpha
features
where,
let's
say:
let's
talk
about
the
production
cluster
right,
so
in
a
production
I'm,
a
production,
cluster
version
111
there
is
not
gonna,
be
nearly
scheduler
right,
gasps,
so,
and
demon
set
parts
are
still
being
scheduled
by
demon,
size
controller.
Yes,
so
yes,
sure.
Yes,
those
critical
parts
which
are
going
to
be
scheduled
by
demons
as
controller
will
never
not
scheduled
because
the
mr.
Cantrell
will
must
will
not
perform
for
you.
D
E
A
E
E
E
A
E
Yes,
a
photo
can
scheduling
way.
We
have
a
Howard
Howard
talk
yesterday
morning,
yeah
and
I.
Think
Oh
are
we
have
a
requirement
on
the
on
the
requirement
and
we
get
word
draft
priority
for
this
one,
a
single.
What
wait
didn't
to
get
the
final
answer
as
how
to
define
the
ATI
object
in
the
core
park
right
yeah.
So,
yes,
the
way
I
propose
the
to
truth
to
object
in
the
document.
Why
is
a
we
I'd
like
to
introduce
two
top?
E
E
A
E
A
E
A
We
introduce
this
here
are
the
all
their
processing
of
these
C
or
D
and
API
objects
or
whatever
what
it.
What
you
can
ask
are
the
objects
or
whatever
should
be
done
by
by
some
of
these
external
controllers
yeah.
So
that's
a
little
bit
of
a
problem,
although
it
could
be
a
solution
as
well
for
now,
basically,
we
can
have
another
component,
something
similar.
D
E
Yes,
that's
excessive,
so
anyway,
I'm
sink
and
roll
hey.
That
also
leads
to
some
other
options,
such
as
the
maybe
he
has
some
jobs,
jobs,
packs
or
something
others
anyway.
I
will
lease
the
option
that
we
can
seem
to
run
now.
So,
yes,
I
sing,
I
will
okay
with
some
suggestion.
For
this
part
and
a
further
comment,
I
think
we
are.
We
are
on
the
same
page
for
this
for
the
requirement
and
what
we
are
going
to
do.
For
this
part,
we
have
a
part
in
list
so.
D
E
A
F
A
So
the
demons
are
controller
basically
used
to
schedule
its
own
parts.
Now
we
are
changing
that
to
make
image
that
controller,
to
get
all
the
faults
that
it
held,
regardless
of
whether
it
can
be
scheduled
right
away
or
not,
and
leave
those
odds
up
to
the
scheduler
to
schedule,
and
this
new
feature
is
marked
as
an
alpha
feature
in
111,
but.
F
A
A
A
The
question
I
I
mean
this
is
not
an
API,
but
it's
more
like
clouds
can
tell
better
but
I
feel
this
is
just
disabled
by
default
and
not
necessarily
alpha,
but
it's
disabled
by
default,
because
we
are
not
so
confident
about
the
feature.
Yet
it's
basically
hasn't
gotten
as
much
mileage.
We
don't
you
haven't
enable
we
don't
know
if
it
could
cause
problem
or
not.
I
guess
that's
why
yeah.
E
I
think
for
the
feature
for
some
feature
we
really
in
both
by
alpha
yeah,
and
this
is
a
disabled
by
default,
and
then
we
will
appropriate
to
beta
I'm
sure
we
get
enough
user
and
get
enough
code
coverage
so
yeah.
So
we
are
going
to
upgrade
to
beta
and
that's
well
yeah
in
bed
how
this
will
be
enabled
by
default.
F
D
A
This
phone
is
ringing
and
ringing
everything
so
I
know
it
was
not
enough
because
it
was
initial
plan
was
to
move
it
to
our
final
110,
but
I
guess
some
issue
happened
and
we
enable
did
it
by
by
removing
basically
or
adding
a
false
and
inside
one
of
our.
If
conditions
that
was
checking
the
flag,
we
will
be
forced
to
disable
the
feature
in
110
and
now
it's
again
in
111.
F
D
F
A
A
B
A
B
A
A
B
A
G
G
If
you
know,
let's
say
it,
saves
a
locatable
and
capacity
for
an
extended
resources.
Let's
say
some
quantity
right
and
I'm
able
to
reschedule
a
pod
and
access
becomes
pending,
and
then,
if
I
delete
some
existing
parts
and
reschedule
the
issue
that
I'm
seeing
is
that
I
don't
see
the
hello
collectible,
decreasing
or
increasing
based
on
the
number
of
odd.
Is
that.
A
Allocatable
doesn't
change
so
on
the
node
side.
You
should
not
see
the
allocatable
getting
changed.
The
allocatable
is
actually
generally
a
fixed
value
as
long
as
the
number
of
those
resources
are
the
same.
For
example,
let's
say
that
you
have
five
GPUs
or
all
those
five
GPS
remained
the
same
on
the
node
side.
As
long
as
you
have
physically,
like
five
GPUs
available
on
under
node,
it
is
the
number
of
those
the
accounting
of
those
happens
inside
the
scheduler
logic.
A
G
G
B
G
G
A
Could
I
mean
no,
it
could
possibly
export
those
as
make
it
right.