►
From YouTube: Kubernetes SIG-Scheduling Weekly Meeting for 20200409
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,
this
meaning
is
recall
this
so
well.
If
you
what
you
said,
you
will
be
public
in
the
in
the
internet,
so
today
we
could
just
discuss
what
we
have
done
in
the
past
few
weeks
and
if
any
interest
a
vehicle,
you
can't
raise
any
time
during
the
meeting.
So
in
terms
of
me,
so
basically
there
are
two
item,
two
major
items
going
on
in
the
last
two
weeks.
Why
is
the
sub
report?
A
So
you
can
check
out
the
iCard
qsq
subtyping
in
the
kubernetes
six
/schedule
plugin
rebel,
so
right
now,
C
is
ready
and
attribute
contribution
is
open
for
all
the
are
the
interesting
parties
who
have
their
custom
plugin
developing
their
companies
as
well
as
you
are
welcome
to
contribute
some
ideas.
You
some
some.
Some
guys
are
interesting
new
idea.
They
can't
get
started
with
the
design.
The
imitation
okay
so
and
also
the
release
plan
has
been
discussed.
A
Regna.
We
still
have
two
vendor
the
two
bananas
slash
Canaries
died
in
the
future.
Maybe
why,
when
Mike's
plan
to
clean
up
the
schedule
dependencies
down,
we
can
remove
the
dependency
on
any
slash,
webinars,
so
yeah
for
the
schedule
fragging
rebel
Mike.
Do
you
want
to
talk
about
something
about
the
proposal
of
collapse?
Crudup
dependencies,
I,
yeah.
B
B
That
needs
to
be
seen.
What
can
we
clean
up
so
some
of
these
changes
will
be
pretty
simple.
Just
move
a
constant
from
one
package
to
another.
Others
might
be
more
involved
when
you
go
to
see
where
those
are
being
used
elsewhere
in
the
code
and
how
we're
using
them.
So
it
seems
to
be
going
well
got
I,
think
for
PR.
So
far,
gonna
get
to
reviewing
the
other
two
or
three
that
I
haven't
yet,
and
a
lot
of
people
seem
to
be
interested
in
helping
out.
That's
just
good.
B
Not
at
this
time,
I
think
I
see
that
there's
a
there
are
some
core
utility
functions
that
we're,
depending
on
that's
one
of
those
things
that
we're
gonna
have
to
look
at
more
deeply,
how
they're
being
used.
What
other
places
are
using
them,
we're
the
only
people,
that's
using
like
one
of
these
core
utility
helpers.
B
D
So
just
have
to
keep
an
eye
out
those
yeah.
From
my
experience.
It's
usually
case-by-case
it's
hard
to
address
this
kind
of
changes
on
a
global
base,
because
it
will
depend
on
the
scale
of
the
changes
that
are
required.
Whether
we
can
share
the
code,
whether
it's
possible
to
duplicate
a
code
in
some
instances,
if
it's
a
small
helper
or
whether
we
actually
need
to
think
about
some
kind
of
a
shared
library
where
we
will
be
extracting
that
code.
Yeah.
A
The
next
thing
is
about
the
preemption
exchange
points,
so
I
justed
a
proposal
in
the
Google
Doc
and
I
receive
a
lot
of
feedback.
So
basically
that
feedback
comes
to
three
possible
solutions.
We
can
go
with
wise
to
use
several
expansion
points
to
describe
each
internal
face
of
the
whole
preemption
book,
though
that
is
first
one.
The
second
one
is
to
use
one
external
point
and
inside
an
extreme
point
we
define
some
internal
private
functions
to
describe
the
whole
workflow
of
the
preemption
and
the
third
one
is
like
Bobby's
that
we
can.
A
We
can
not
enforce
the
semantics
of
that
X
point
to
be
preemption.
We
can
just
say
you
can
implement
any
logic
as
the
post
of
filter,
which
is
just
invoked
after
the
filter.
So
basically
you
can
say
I
don't
want
preemption
right,
I
just
want
to
get
notified
when
the
filter
failed,
so
I
can
take
some
other
attractions
like
invoke
the
cluster
autoscaler
to
provision
a
new
node
or
just
invoke
the
HPA.
A
Sorry,
not
HPA
vp,
a
vertical
part
autoscaler
to
reduce
the
paths,
requests
or
limits
or
sort
of
so
that
option
is
the
third
one
which
doesn't
you
foster,
semantics
of
preemption,
so
the
latter
two
options
actually
need.
Some
pre
factoring
work
to
make
that
happen,
because,
right
now
those
kind
of
code
are
internally
wired
to
some
other
internal
data
structures
which
we
don't
want
to
expose.
A
So
it
has
to
call
the
theaters
and
also
some
filter
parking
relies
on
pre-filter,
which
is
just
which
is
a
pre
calculation
on
some
kind
of
metadata
and
some
cash
to
make
the
further
calculation
faster.
So
it
has
a
lot
also
to
call
the
pre-filter
clunking
as
well,
and
also
some
some
logic
are,
depending
on
some
scheduling
queue.
So
a
lot
of
internal
stuff
wired
together.
So
so
we
have
to
separate
those
dependencies
first.
Otherwise
we
cannot
get
to
a
conclusion
which
approach
we
are
came
back
to
it
eventually
go
with.
A
So
that
is
one
thing:
I'm
going
to
send
out
a
key,
be
sure
to
a
brainstorm.
I.
E
B
E
Don't
really
need
to
try
preemption
with
or
a
scaler
behavior
and
whatnot.
We
can
just
say
okay,
this
is
the
preemption
object
at
a
default.
Scheduler
supports-
and
here
are
a
couple
of
you
know
that
you
can
implement
to
alter
its
behavior
I
would
have
liked
if
we
can
make
it
more
generic,
but
I
think
it's
completely
fine.
If
we
get
to
the
conclusion
where
this
is
not
feasible
or
it
makes
our
life
super
complicated.
Yes,.
A
E
Import
a
framework,
the
framework
package
and
the
framework
package
itself
only
you
know
only
depends
on
itself
doesn't
have
any
example
is
anything
that
makes
sense
they
shouldn't
be
extended,
like
should
reach
out
and
do
anything
outside
and
anything
that
would
require,
for
example,
the
favor
to
dependent
anything
outside,
for
example,
a
snapshot.
It
has
to
be
behind
the
interface,
so
this
framework
defines
an
interface
and
then
at
instantiation
time
the
logic
that
instantiate
the
framework
would
provide
a
specific
documentation
by
that
interface,
and
so
there
is
no
outward
dependencies.
E
I
work
around
providing
pot
info
and,
and
so
they're
like
we
had
not
info
outside,
we
had
the
listservs
outside
and
all
of
those
concepts
now
can
easily
be
moved
inside.
So
this
is
what
I
did
and
inside
the
framework
we
will
have
the
framework
interface.
We
will
have
the
types,
the
Lister's
and
then
as
I'll
go
suggested.
E
Basically,
clean
up
first
mode
info:
there
is
a
lot
of
logic
there
that
is
I
think
is
obsolete.
The
needs
to
be
cleaned
up
and
not
employ
itself
now
is.
Is
the
lid
is
the
type
that's
being
returned
when
you
access
the
pods
on
a
node
through
node
info,
so
in
a
in
a
follow
up,
be
ours?
We're
going
to
include
more
information
in
that
word
info
that
we
can
pre
calculate
once
basically
like,
like
selectors
and
whatnot,
that
filters
and
other
plugins
can
rely
on
without
completing
at
en
the
ending.
A
E
C
A
A
You
and
I
think
that's.
Another
thing
is
that
out
also
some
is
some
to
peers
to
improve
the
performance
of
part
to
parish,
spread
the
first
one
is
to
trade,
the
case
that
the
part
which
wants
to
spread
the
past
on
to
the
house
as
a
special
case,
because
in
that
case,
if
it
has
a
lot
of
nails,
the
internal
data
structure
wasn't
optimal
so
that
there
was
some
optimization
on
that.
The
second
one
is
to
some
somehow
internally
use
the
use
the
atomic
package
against
a
point
instead
of
using
explicit,
lock
onto
integers.
A
C
Yes
sounds
good
yeah.
One
thing
to
know
there
is
that
we
got
to
a
point
where
we're
score
is
faster
than
scoring
is
faster
than
filtering.
We're
using
you
know,
call
spread
and
he's
because
of
of
the
way
how
we
calculate
this
call
in
which
we
don't
depend
on
our
notes.
Whereas
for
for
a
filter
we
will
depend
on
other
notes,
so
we
need
the
pre-processing
earlier.
So
we
need
to
keep
this
data
structure
with
the
map,
which
is
which
is
slower
so
yeah
I.
C
C
A
E
A
F
Yeah
I
want
to
talk
about
scissor
lazy
evaluation
about
your
resource
quarter.
I
had
some
talk
of
the
idol
and
I
want
sure,
as
some
culture
to
describes
the
namespace
a
hard
Tigger
and
I
want
you
and
says
the
code
has
lemon
kind
of
request
of
the
juicer
and
a
true
evaluate
rating
is
a
scheduler
and
I
think
if
we
have
this
Y
in
the
scheduler,
we
will
true.
A
Can
describe
something
because
Alex
talked
to
me
offline,
so
basically
there
there
was
a
requirement
I
think
it's
long
discussed
in
the
community,
which
is
there.
They
want
a
soft
resource
coder
which
doesn't
take
into
effect
during
the
admission
time,
which
is
upon
the
API
object
creation,
for
example,
part
of
creation.
So
they
want
to
the
resource
kata
to
be
evaluated
later
during
the
school
scheduling
time
so
that
they
can.
For
example,
two
teams
share
the
culture
of.
A
Maybe
one
team
use
a
lot
and
the
other
is
less
all
the
vice
versa,
so
that
is
one
I
think
which
is
maybe
a
common
requirements
during
the
batch
workers.
They
want
to
submit
as
a
series
of
what
close,
but
they
don't
what
close
doesn't
need
to
be
executed
at
the
same
time,
they
just
want
to
put
the
pods
into
the
queue
and
internally,
scattering
can
adjust
that
called
her
assigned
to
each
of
its
operation.
A
Namespace
or
each
soft
attain
and
balanced
way
or
one
tenant
can
borrow
some
resource
Cola
for
another
sort
of
sort
of
like
this.
This
kind
of
thing
so
I
should
have
an
issue.
I.
Remember
some
someone
open
one
year
or
two
years
ago,
I
think
Alex
and
his
company
has
some
internal
practice
on
that.
So
it
will
be
good.
It
can
be
such
abstracting
to
a
cat
and
go
with
this
implementation
for
some
Heiser.