►
From YouTube: Kubernetes SIG Scheduling weekly meeting - 2018-12-06
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
A
Basically,
I'll
put
that
issue
like
most
of
the
cars
are
already
merged
for
the
case
that
a
pod,
which
is
tried
frequently
with
the
same
priority
compared
to
other
parts
in
the
queue,
is
not
a
scheduler
pulling
up
on
the
schedule
of
all
part
or
parts,
get
blocked
the
head
of
the
queue
so
that
that
one
is,
as
we
are
working
on
a
couple.
Other
improvements
for
for
the
scheduling
queue.
One
is
the
back
off
mechanism
for
the
scheduling
queue.
A
B
A
It's
had
a
meaningful
change.
Both
of
these
PRS
are
out
there.
We
are
reviewing
them
and
hopefully
we
can
merge
them
soon.
There
was
a
performance
regression
and
this
scheduler
turned
out
to
be
due
to
a
change
that
we
made
in
priority
functions,
which
was
actually
more
of
a
refactoring,
but
surprisingly,
it
cost
performance
degradation.
When
we
looked
further
in
the
code,
we
realized
that
the
previous
code
was
running
almost
all
the
priority
functions
in
parallel,
but
after
we
refactored
the
new
code
was
running
the
old-style
priority
functions,
which
was
not
in
that
MapReduce
format.
A
First
and
then,
once
those
finished
it
would
start
a
new
style
which
is
in
MapReduce
format.
So
I
changed
this
back
to
the
previous
style
or
previous
parallel
algorithm
and
it
seems
to
address
the
performance
regression.
These
are
some
of
the
updates
today
and
that
I
wanted
to
give
you
regarding
things
that
happen
in
a
past
week,
or
so
there
are
items
that
I
would
like
to
talk
about.
Since
some
of
the
folks
are
here
today,
one
is
actually
Robbie.
A
C
A
You
know
also,
we
were
thinking
about
promoting
the
scheduler
as
a
standard
component.
One
of
the
one
of
the
features
that
was
lacking
was
the
fact
that
this
scheduler
today
does
not
have
a
simulation
mechanism
to
you
or
other
folks
in
your
organization.
Have
any
plans
to
add
that
Mike
Arizona
scheduler
at.
C
A
So
you
know
one
of
the
things
that
I
actually
don't
have,
given
that
your
organization
has
like
major
changes.
Recently,
I,
don't
know
if
how
is
your
it
roll
down
or
whether
you
guys
are
still
actively
working
on
projects
that
you
use
to
work
or
not?
So
that's
an
area
that
we
don't
know,
but
if
you
have
the
time
you
know
that
would
certainly
be
useful.
I
think
that's
something
that
we
either
I
mean
the
open-source
community.
We
should
either
find
other
folks
who
are
willing
to.
A
C
A
C
At
this
point
of
time,
people
have
formed
alternatives,
not
using
scheduler
as
such,
but
they
have
found
out
Ignotus
to
this
functionality,
but
if
you
think
that
there
is
a
need
in
the
community
as
such,
the
only
thing
remaining
in
this
feature,
especially,
is,
is
trying
to
find
out
the
performance
aspects
like
how,
if
they
include
the
player,
if
we
include
this
priority
function
by
default,
is
there
going
to
be
any
performance
degradation?
That's
something
that
we
need
to
evaluate
right.
C
A
So
that
was
basically
the
main
blocker
for
moving
this
to
beta.
We
didn't
have
the
answer
to
that
question
on.
We
felt
that
we
felt
that
this
feature
might
slow
down
the
scheduler.
So
I
don't
know
I
this.
This
is
something
that
I
believe
you
proposed
I,
don't
have
any
strong
use
cases
for
it
at
the
moment,
but
if
it's
something
that
you
want
us
to
pursue,
we
will
be
happy
to
continue
working
on
it
with
the
community
of
course,
and
find
out
if
this
is
performant
enough
for
now,
yeah.
A
Sure,
okay,
thank
you
very
much.
Riley
wait!
You
were
here
so
I
saw
that
you
I
guess
you
commented
on
one
of
these
features
that
you're
pursuing,
which
is
like
supporting,
inter
pot
affinity
to
more
than
one
pot.
So
you
mentioned
that.
Apparently
you
may
not
have
time
to
work
on
it
for
114.
Is
that
right
or
why
I
misunderstood
no.
D
A
D
A
D
A
Like
it
was
almost
there,
even
in
113,
but
since
113
was
supposed
to
be
a
stability
release,
we
didn't
want
to
make
it
push
it
so,
but
I
agree
with
you,
I
feel
I
feel
like
it's.
The
supporting
means,
apart
effect
to
more
than
one
part,
is
almost
done.
Okay,
so
this
is
like
mixed
money
max
mean
whatever
yeah
I
I
saw
that
it's.
The
comment
is
in
the
right
place
on
the
spreadsheet
yeah.
D
A
Sounds
good,
okay,
I
guess
these
are
all
the
updates
for
now
that
I
have
or
I
wanted
to
hear.
As
you
know,
next
week
is
kill,
Khan
of
North
America
I
want
to
be
here.
I
will
be
attending
coupon.
If
you
are
attending,
I
will
see
you
there
in
person
and
we
are
not
going
to
have
any
scheduling
meeting
next
week
around
the
same
time
actually
like
like
4:00
or
4:30
p.m.
Pacific
time,
there
is
gonna,
be
six
scheduling.
A
D
Don't
got
question
but
want
to
kind
of
promote
my
statute
a
little
bit
so
on
the
second
day.
Second
day
should
be
wednesday
of
kill
car
at
the
loss
of
time.
Slot
I
will
be
giving
a
session
about
how
to
extend
cabin
area
load
balancer
in
different
cloud
providers.
So
it's
not
the
exactly
related
with
scheduling,
but
I
will
mention
some
concept
and
the
knowledge
which
I
learned
from
sting
scheduling.
It's
more
relatable
Signet
what
else
T?
So
you
said
it's
the
second
day
of
kill
con.
D
B
Hi
I
just
wanted
to
introduce
my
son.
It's
the
first
time
I've
attended
this
meeting.
My
name
is
Diana
Royale.
Actually,
I
should
turn
my
camera
on.
So
you
can
see
it
me.
Okay,
I
I've,
been
working
behind
the
scenes.
Essentially
I
think
MA
has
been
involved
in
this
working
group
and
I
thought.
I'd
start
participating
a
little
bit
just
to
get
familiar
with
the
team
and
the
work
that's
being
done
as
well.
B
D
A
A
E
Definitely
yeah
so
far
as
I
sent
out
an
email
last
week,
so
we
ran
next
round
of
tests.
After
the
you
know,
the
gang
scheduling
and
the
Mac's
part
and
all
those
new
functionality
so
yeah.
So
the
trend
pretty
much
stays
the
same.
So
as
long
as
resource
requirements
are
uniform
across
jobs,
we
see
pretty
substantial
throughput
for
normal
parts,
but
the
the
part
of
feminity
went
down.
So
you
guys
did
lot
of
optimization
actually
as
far
as
the
111,
so
our
our
part
affinity
is
essentially
implemented
exactly
like
the
default
scheduler.
E
Currently,
the
one
part
at
a
time,
because
there
is
a
better
way
of
doing
it
in
network
flow
graph
approach,
but
we
didn't
so.
This
is
based
on
the
theses.
Actually,
you
know
and
multi
published
in
from
University
of
Cambridge,
and
we
ran
into
a
lot
of
problems.
So
we
cannot
discontinue
that.
So
this
is
kind
of
a
stopgap
thing
for
us,
so
we
will
see
if
we
can
do
some
kind
of
a
short-term
kind
of
combination
so
that
we
stay
pretty
much
the
same
level
there.
E
The
default
scheduler
is,
but
our
goal
is
essentially
eventually
to
go
down
the
path
to
do.
The
actual
flow
graph
base
power
definitely
useful
so
and
then,
based
on
our
discussion.
Bobby
Esau
am
in
the
process
of
putting
together
all
the
documentation
stuff
and
all
that
to
essentially
include
the
alternate
scheduler
Poseidon
for
moment
as
the
alternate
scheduler
and
bunch
of
things
introductions
and
way,
which,
when
would
you
want
to
use
it
and
that
key
advantages
and
what
stage
yeah
so
just
kind
of
a
brief
yeah
yeah
to.
A
Just
give
a
little
bit
more
context
to
people
who
are
not
fully
aware
of
this
Poseidon,
as
some
of
you
folks
know,
is
a
firmaments.
Scheduler
is
based
on
time
on
schedule.
Actually,
it
has
a
very,
very
high
throughput
compared
to
our
currently
scheduled
air
for
for
scheduling
a
large
number
of
parts
with
similar
scheduling
properties.
It's
not
much
better
than
our
current
scheduler
in
all
scenarios.
So,
for
example,
recent
resource
at
debug
sent
out
shows
that
it's
not
that
great.
Compared
to
our
current
scheduler
for
scheduling
parts,
we
definitely
not
the
affinity.
A
This
is
what
he
also
mentioned.
I
just
mentioned
basically,
and
but
our
plan
is
to
introduce
this
as
an
alternative
scheduler
for
people
who
may
need
to
have
very
high
throughput
for
the
schedule.
Of
course,
it's
also
not
100%
compatible
with
all
the
API.
So
maybe,
if
we
use
certain
properties
of
parts,
it
may
not
work
well,
one
area
that
is
identified
is
actually
volume
binding
or
dynamic
volume
binding.
That
sedan
doesn't
work.
Yeah.
E
So
we'll
have
another
thing
which
we
are
doing
Bobby
at,
as
you
mentioned,
we
are
going
to
put
together
a
cable
act,
so
that
was
kind
of
say,
okay,
so
these
are
the
features
which
default
scheduler
has
and,
for
example,
dynamic
volume
binding.
So
we
cannot
list
out
each
one,
one
of
them
and
say
yes
or
no
in
front
exactly.
A
E
A
Any
other
question
or
comments.
You
know
I've
said
many
times
that
if,
if
your
peers
need
my
review,
please
let
me
know
on
slack
or
somewhere
else
or
assign
them
directly
to
me.
Otherwise
chances
or
I
won't
be
seen
or
I
won't
see
them.
Please
continue
doing
so.
If
you
want
my
attention,
all
right
looks
like
there
is
no
other
or
comment,
so
you
get
ten
minutes
back
I.
Thank
you
very
much.
I
hope
to
see
see
you
folks
next
week
at
cube,
run
by
you,
Thanks.