►
From YouTube: Kubernetes SIG Scheduling meeting - 2018-09-20
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,
so
let's
just
artistic
scheduling,
meeting
and
let's
start
with
or
without
planning
and
agenda
for
for
the
next
release
cycle,
so
I
can
give
you
a
little
bit
of
update
about
some
of
this
stuff.
That
I'm
aware
one
with
regards
to
equivalents.
Cash
actually
Harry
has
written
a
document
who
he
shared
a
document
actually
with
me
and
I
I
commented
a
document
and
did
minor
changes
to
the
document
he's
going
to
send
it
soon
to
other
folks
I'm
into
this
to
the
seed,
so
that
other
people
can
take
a
look
as
well.
A
So
the
idea
here
is
that
we
are
gonna
change
the
equivalents
cash
design
so
that
it's
more
appropriate
basically
for
a
four
clusters
that
have
a
lot
of
equivalent
spots
that
cannot
be
scheduled
and
instead
of
scheduled
trying
to
schedule
all
of
these
parts.
Again
and
again,
it
tries
one
of
these
pods
and
if
the
pot
turns
out
to
be
under
schedulable,
the
rest
of
them
won't
be
tried.
So
that's
that's
the
idea.
Behind
the
new
equivalents
cash
design,
we
are
hoping
that
we
can
get
some
performance
improvement
out
of
this
new
design.
A
In
of
course,
select
scenarios
not
in
all
the
parts
of
scenarios.
Of
course,
the
previous
design
was
expected
to
help
us
in
more
or
less
a
larger
or
broader
number
of
scenarios,
but
it
turns
out
that
our
scheduler
predicates
are
already
fast
enough,
or
most
of
them
are
fast
enough,
so
that
equivalents
catch
didn't
make
a
whole
lot
of
sense.
So,
with
the
new
design,
we
are
hoping
that
we
can
address
those
issues.
A
B
Yeah
we
house
the
roll
Ronna
really
hope
out
the
KP
I
think
their
API,
so
API
or
the
API
is
here
Jana.
We
also
have
a
agreement
on
the
transaction
part.
There's
some
corner
kids
away.
I
need
to
to
thinking
about
such
as
details
of
accounts,
getting
the
the
task,
or
maybe
the
results
maybe
hold
farewell.
So
we
meet
I
will
time
out,
or
maybe
some
other
things,
and
there
are
also
have
some
comments
about
the
lifecycle
of
a
pod
group.
Yeah
I
think
for
this
task
for
the
first
version
of
account
scheduling.
A
Sounds
great
yeah.
Thank
you
very
much
for
your
efforts.
Okay,
so
you
know
I've
already
put
it
put
this
other
item
in
our
spreadsheet
as
well.
We
are
planning
to
implement
a
couple
of,
or
at
least
one
probably
two
extension
points
and
the
current
scheduler
as
a
as
a
part
of
the
features
that
we
are.
We
would
like
to
build
from
the
scheduling
framework.
A
So,
with
these
extension
points,
we
will
also
design
the
interface
for
the
plugins
in
process
plugins
and
we're
hoping
that
after
we
build
this,
we
will
be
able
to
make
some
of
these
existing
hard
coded
pieces
of
scheduling,
particularly
dynamic,
dynamic
resource
allocation
or
done
abusers
binding
convert
those
to
a
to
a
plugin.
So
that's
our
hope.
A
We
will
see
how
much
we
can
achieve
in
the
next
quarter,
and
this
quarter,
as
many
of
you
know,
is
a
shorter
quarter
as
much
further
given
the
New
Year
holidays,
the
code
freeze
and
everything
which
usually
happens
earlier
there
than
before.
So
we
will
see
how
much
we
can
get
in
before
the
next
code.
Freeze
we,
we
are
also
finalizing
the
design
of
party
scheduling
policies.
You
may
be
aware
that
we
were
working
with
this
with
the
policy
workgroup
and
we
are
finalizing
the
design
of
the
scheduling
policies.
A
A
Okay
and
yeah,
we
would
like
to
move
taint
based
eviction
to
beta
I.
Don't
think
that
there
has
been
want
to
update
our
with
regards
to
that
Ravi
or
way,
if
any
of
you
guys
have
updates
on
that
I
doubt
it.
But
if
you
have
any
updates
on
that,
you
guys
started
working
on
it
yet
or
not.
Our
fraud
is
actually
here.
I.
C
A
That's
like
that's
what
I
guessed
as
well.
Okay,
we
should
should
start
basically
planning
on
that
I'm.
Actually,
given
Ravi
has,
let's
see
actually
thought
was
also
working
on
another
okay
he's
not
okay,
I
was
going
to
say
if
Ravi
has
many
items,
we
can
give
this
to
you
but
looks
like
he
doesn't
so
one
other
thing
that
we
would
like
to
do
is
to
deprecate
critical
part
annotation
and
that's,
hopefully,
not
a
large
effort.
You
can
deprecated
that
in
1:13
and
get
rid
of
it
at
some
point
in
the
future.
A
I
know
that
Ray
has
been
working
on
supporting
interpret
affinity
to
more
than
one
heart.
So
the
idea
here
is
that
currently,
basically
in
the
current
current
implementation
of
heart
enter
pod
affinity,
we
expect
one
part
to
satisfy
all
the
affinity
terms,
so
an
affinity
term,
an
affinity
rule
wasting
disorder.
The
affinity
on
a
pod
can
have
multiple
terms,
so
these
terms
or
are
described
as
a
label
select
or
a
name
space,
and
then
some
topology,
and
we
expected
multiple
of
these
terms
to
be
satisfied
by
one
part.
A
The
reason
that
we
decided
to
do
this
initially
was
the
fact
that,
in
a
pod,
affinity
was
very
complex
and
very
time-consuming,
so
it
was,
as
was
very
slow,
and
if
you
wanted
to
support
multiple
part
matching,
it
would
have
become
even
much
much
slower
with
a
new
implementation.
However,
we
think
that
we
can
support
it
much
more
easily,
basically
support
multiple
parts,
matching
the
inter
pot
affinity
of
an
incoming
pot,
and
we
actually
made
the
changes.
C
D
C
I
want
to
reduce
the
time,
complexity
and
the
space
complexity
by
using
an
unsigned
into
64
to
record
which
index
of
the
affinity
match
some
there
on
a
topology
key,
so
so
that
so
finally
I
will
look.
Cap
I
will
kind
of
use
a
bitwise
or
calculation
order
on
individual
items,
so
so
that
is,
could
reduced
complexity
so
I'm
not
sure
if
this
limitation
makes
sense
to
you
actually.
A
I
was
looking
at
your
pure
and
I.
Wasn't
so
sure
why
why
we
needed
that?
But
anyway,
you
know
I
to
be
honest
with
you,
I,
don't
think
that
anyone
who
will
put
like
a
very
large
number
of
key
self
energy
charms
and
the
reason
that
I'm
saying
that
is
that
it
becomes
very
soon
out
of
hand
or
odd
of
the
capacity
of
humans
to
understand
what
is
going
on.
If
you
put
like
I,
don't
know
a
thousand
affinity
terms
on
a
single
part,
then
nobody
almost
nobody
understands
what.
A
E
A
Be
really
worried
about
having
a
large
number
of
them
anyway.
I
don't
want
to
dive
into
the
details
because
a
lot
of
people
here
talking
about,
but
we
will
continue
the
discussion
on
the
PR
okay.
So
there
is
one
item
which
is
not
assigned
to
anyone.
Yet
actually
that
item
is
is
building
a
new
new
test
library
for
for
building
scheduler
benchmarks.
A
We
currently
have
some
scheduler
benchmarks,
which
are
part
of
our
integration
tests,
and
these
these
benchmarks
are
rather
hard
coded
and
rigid
I
would
call
you
know
they
are
very
specific
and
they
are
written
in
a
very
specific
way
and
the
number
of
the
number
of
nodes
and
parts
and
everything
in
those
benchmarks
is
very
fixed
but
much
fixed
on
the
hard
coded.
So
one
one
thing
that
I
was
thinking
about
was
was
basically
creating
a
library
that
would
allow
us
to
build
these
benchmarks
more
easily.
A
One
particular
aspect
that
I
would
like
to
see.
Those
for
those
benchmarks
is
to
build
new
urban
new
benchmarks
for
our
various
features
that
we
are
adding
to
the
schedule
to
measure
their
impact
on
performance
and
our
current
platform
and
the
libraries
that
we
have
do
not
really
support
adding
benchmarks
easily.
So
that's
one
thing
that
we
need
to
do.
I
sometimes
feel
like
this
is
something
I.
I
need
to
do
myself
so
building
the
integration
test
library,
because
mostly
because
I
have
something
in
my
head
and
I
cannot
really
put
it
on
paper.
A
B
A
A
A
B
B
Yes,
I
think
pata
group,
whether
it's
a
not
complainant
for
new
rec
or
new
players.
So
a
nice
has
really
tailed
to
the
Badger
scatterer
yeah.
Yes,
in
a
way
under
folk
who
Betty
as
a
automation
before
this
maybe
give
us
some
attention
that
chooses
gather
of
him
work,
a
full
patch
power,
yeah,
so
I
think
yeah
so
yeah,
for
when
scheduling
work
is
ready
and
we
can
consider
how
to
more
to
counter
work
together
for
one
for
one
scatterer
right.
So.
A
You
know
you
know
what
our
plan
for
building
this
the
framework
so
case
we're
gradually
building
those
features
inside
our
trenches,
scheduler
and,
as
you
know,
we
have
already
some
plan
for
building
a
couple
of
those
extension
points
in
an
experience.
Yes,
hopefully
in
the
next
release
you
will
see,
but
but
you
know,
we
can
prioritize
the
features
that
you
need
for
for
batch
scheduling,
for
example
yeah
yeah.
If
you
need
certain
features,
we
can
prioritize
building
those
first
before
them
from
with
respect
to
Q
party
I.
E
B
I
just
follow
our
yes,
some
components
in
our
product,
as
with
the
D,
so
I'm
I'm
also
I'm,
also,
okay,
with
some
other
name.
Okay,
our.
A
A
A
I
was
I
would
say
the
same
thing,
so
I've
not
necessary
because
it's
it
sounds
cool,
but
it's
a
very
well-known
keyboard.
Really
everybody
knows
what
I
I
was
thinking
that
everybody
knows
why
gang
scheduling?
Yes,
so
we
should
he
thank
you,
but
but
it
turns
out
that
actually
most
people
don't
know
what
gang
scheduling
is
and
people
given
the
connotation.
People
may
think
that
it's
something
related
to
Yanks
and
I
I
was
very
raised
quite
and
people
thought
that
way.
But
anyway,
so
they
didn't
really
like
the
name
of
gang.
E
F
B
F
E
B
C
A
B
Okay,
so
we
can
trace
their
history
there.
There
are
already
some
you
see
and
some
ID
k
here.
So
yes,
yes,
yeah,
okay!
This
is
for
the
kebab
feature
and
a
new
batch,
and
there
are
also
the
other
stuff.
In
my
hand,
is
some
energy
and
the
tasks
here
yeah
one
I'm,
Trent
Hauser,
is
about
demo
site
upgrade.
Then.
A
B
There
is
a
common
City
Council
in
janitor,
and
it
has
great
didn't
chest
which
has
a
number
of
10th,
but
for
the
following
key
Messiah.
There,
okay
was
on
default,
the
toleration
yeah,
so
the
no
memorize
are
different
between
the
I
pray
running
and
and
the
entrances
excitation.
So
so
the
tattletale
here
actually
yeah
I
will
cover
this
I
will
cover
this
together.
Just
today,
I
think
just
to
talk
with
the
Jeannette.
They
would
like
to
just
remove
the
the
test.
I'll
pause,
the
pant
I
see
yes,
because
the
default.
B
A
B
A
A
E
A
Parts
getting
scheduled
on
a
master
note
in
some
of
our
test
environments
and
highly
I,
really
think
that
it's
probably
the
the
fact
that
master
know
does
not
change
it
properly
in
those
test
environments.
What
do
you
need
to?
We
need
to
actually
verify
that,
but
I
must
I
think
most
probably
the
case.
Yes,.
B
I
just
see
the
comments
about
their
power.
I
didn't
go
to
the
detail,
yes,
I
think
I
will
I
will
pull
sometime
today
on
these
issues,
because
the
other
one
years
or
the
rule
cows
there
and
the
fix
is.
We
already
know
how
to
fix
that.
That's
we're
gonna
be
a
simple
to
easier
to
do
that
and
then
I
will
pull
some
time
on
this
one
yeah,
okay,
think
I
was
up
law.
We
can
to
see
what
happened.
Yes
sounds
great.
Thank
you
yeah,
my
pressure.
B
A
B
E
A
B
B
F
F
A
A
So
in
that
section
it
mentions
some
of
the
items
there.
I
search
actually
tells
one.
We
need
to
delve
a
thingy
in
the
first
phase,
that's
there,
but
there
is
one
kind
of
like
missing
piece
and
in
this
design
and
that's
finding
the
interface
for
for
for
the
plugins.
So
basically
the
framework
will
always
call
the
plugins
expecting
a
certain.
You
know
face
right,
so
you're
always
passes
I,
don't
know
scheduler
cash,
maybe
on
some
other
part
and
some
some
of
this
information.
A
So
we
haven't
figured
out
what
that
what
does
interface
alright
and
that's
an
important
piece
that
we
need
to
figure
out.
So
that's
why
I
haven't
actually
much
further
to
ask
people
to
help
us
build
that
built
a
framework,
but
what
I
feel
we
should
be
I
should
probably
do
with
the
help
of
community
we've
finalized
that
interface,
maybe
in
a
pure
or
something,
and
then
we
can
open
up
to
the
next
phases
of
building
plug
in
points
and
extension
points
for
her
updating
the
cache
of
the
scheduler.
A
A
F
A
E
I'm
sorry
I
have
a
couple
of
things
that
I
wanted
to
talk
about.
First
thing
is:
there
are
two
features
that
are
in
alpha
stage.
One
is
a
priority
function
which
actually
takes
limits
on
the
pod,
which
was
Alfred
by
sort,
which
was
an
alpha
stage
that
was
created
by
aware
some
time
ago.
I
I
do
not
know
if
we
could
push
it
to
beta
in
the
next
release
in
1.13
is.
A
I
think
so,
I
I
think
it
makes
sense
to
have
that
at
that
priority
function.
I
I,
don't
have
any
objections
really,
of
course,
we
must
ensure
that
we
have
proper
tests
and
everything
for
it
and
we
should
probably
give
it
a
lower
weight
compared
to
many
other
priority
functions
such
as
affinity,
anti
FN,
ad
and
I.
Don't
know
you
should
actually
go
and
look
at
some
of
these
or
maybe
I
misspoke
items
on
some
of
this
stuff,
so
you
should
give
it
proper
weight,
but
otherwise,
yes,
I
am
fine
with
that
yeah.
E
D
A
E
A
A
Yeah
I
mean
for
the
purpose
of
that
particular
priority
function.
I
think
we
need
to
ensure
that
the
performance
is
there,
that's
one
thing
and
actually,
if
you
think
you
can
build
some
benchmarks
with
the
existing
with
the
existing
code
that
we
have,
that
will
be
great,
otherwise
I,
don't
know.
We
should
ensure
that
performance
is
direct
by
other
means,
maybe
building
some
other
tests
similar
to
whatever
we
have,
which
is
kind
of
like
see.
E
C
Just
wants
minor
update
for
me
is
my
boss
is
pushing
me
to
do
some
documentation.
Blogging,
prepare
some
materials
for
or
for
scattering
workshop.
So
one
one
point
thing:
I'm
trying
to
do
is
to
kind
of
summarize
all
our
stuff
related
with
how
to
customize
the
schedules.
Our
ugliest
may
be
all
the
ways
to
customer
that
customers
bad
and
put
them
together
with
some
samples
like
that
yeah
to
customize
the
scheduler
yeah,
both,
for
example,
using
the
scheduler
API
and
also
using
the
extender
mechanism
and
putting
things
together.
Yeah
sounds
good.
D
A
Yeah,
whichever
works
for
you
really
so
as
long
as
YouTube
folks
know
exactly
how
you
want
to
split
the
work,
it
should
be
fine
with
me
and
we
should
basically
ensure
that
we
can
keep.
We
can
make
the
deadlines
basically
so
right
now,
the
dull
human,
the
design
duck,
is
not
finalized,
of
course,
so
it
will
still
take
some
time
until
we
or
fully
finalizing
it.
I
think
we
are
almost
there,
but
you
know
how
things
work
I
need
to
ensure
that
everybody
agrees
and
everything.
So
it
usually
takes
days.
A
A
There
is
no
harm
in
talking
to
him
earlier,
actually
reach
out
to
him
and
see
what
his
plans
and
schedule
for
for
building
discard
the
party
scheduling
policies
and
if
he
is
open
to
you,
know
working
with
you
on
building
those,
and
hopefully
there
is
I
named
pretty
sure
there
is
a
lot
of
work
for
more
than
one
person.
Okay,.