►
From YouTube: Kubernetes SIG Scheduling Weekly Meeting for 20230223
Description
Kubernetes SIG Scheduling Weekly Meeting 2023-02-23T18:00:07Z
A
A
B
B
Okay,
yeah
yeah.
This
is
vs
code.
So
last
time
we
met,
I
was
telling
about
this
need
to
have
a
plug-in
into
the
scheduler,
which
would
do
topology
scoring
and
as
a
recap
of
that,
the
idea
was
to
you
know
basically
do
resource
scoring
similar
to
what
no
resource
feed
is
doing,
but
now
for
topologies.
B
So
suppose
you
have
like
four
topologies,
where
there
are
nodes
and
in
this
example
here
one
of
the
topologies
has
two
machines,
and
assuming
that
you
know
the
machines
are
relatively
unloaded
or
have
a
similar
note.
Then
this
bottom
topology
here
would
have
like
significantly
more
resources
than
the
others.
B
So
the
intention
with
this
new
plugin
was
that
to
be
able
to
give
a
score
for
the
entire
topology,
so
that
the
best
topology
would
nodes
would
get
like
a
hundred
each
and
then
the
worst
topology
could
be
like
this
would
get
a
zero,
and
you
know
those
topologies
that
that
would
have
resources
somewhere
in
between
the
worst
and
the
best
would
get
something
linearly
as
they
score.
So
that
was
the
the
idea
and
I
was
suggested
to
do
an
entirely
new
plugin.
B
Instead
of
fiddling
with
the
no
resources
fit
itself,
so
that's
where
it
started
from,
and
the
thing
is
that
to
do
this
kind
of
topology
scoring
I
would
actually
need
the
score
values
from
the
node
resource
fit,
which
is
doing
things
quite
nicely,
and
it's
nicely
configurable
and
so
forth.
So
I
was
thinking
that
okay,
if
I,
do
a
new
plugin.
What
are
my
Alternatives
do
I
like
copy
the
code
of
no
resources,
fit
that
wouldn't
make
much
sense.
B
I
don't
want
to
do
a
copy
paste
code,
so
is
that
I
first
created
an
implementation
with
composition,
and
that
was
fairly
fairly
simple
but
and
actually
I
could
I
could
show
that
what
it
looks
like
it
was
a
little
84
lines
for
the
basically
the
wrapper
part
of
it
and
the
main
idea
there
was
just
to
have
a
score
plugging
interface
in
a
new
plugin.
This
composition
example,
and
basically
just
instantiate
a
node
resources
fit
into
there
and
place
it
there,
and
at
that
point
it's
already
like
ready
pretty
much.
B
It
can
run
if
you
just
have
the
rest
of
the
boilerplate
there,
and
so
it
did,
and
that
was
kind
of
fine
and
to
get
the
basically
the
score
back
to
this
new
plugin.
All
I
had
to
do
was
to
basically
add
the
score
extension
return,
something
else
than
nil
and
the
framework
delivers.
The
calculations
from
this
included.
B
Node
resources
fit
to
the
normal
last
score
function
then,
and
I
can
do
my
magic,
then
calculating
those
topology
scores,
which
is
not
that
interesting,
probably
for
anybody,
but
but
this
was
my
first
take
at
you
know
how
to
reduce
this
existing
plugin
and
to
get
and
in
terms
of
how
much
code
we
went
there
not
much
so
it
became
pretty
look
pretty
nice,
okay,
at
least
in
my
opinion.
What
do
you
guys
think
not
big.
A
B
A
A
B
I
want
the
full
score,
so
the
drawback
with
this
design
here
is
that
it
actually
runs
the
same
node
resources
fit
calculations
twice,
so
that's
not
ideal.
It
could
have
been
ideal
to
get
just
the
score
from
node
resources
fit
and
continue
from
there.
So
I
didn't
like
this
too
much
and
the
other.
Basically,
let's
say
drawback.
That
I
saw
was
related
to
the
actual
configuring
of
this
thing,
because
I
need
to
have
the
node
resources
fit
arguments.
B
Unfortunately,
this
is
not
updating
properly,
but
let's
see
if
it
now
it
does.
Okay,
so
I
came
up
with
this
so,
and
this
is
by
the
way
in
this
it's
a
little
bit
more
complicated
than
is
necessary,
so
only
the
top
right
corner
is
actually
necessary
for
my
use
case,
but
for
the
sake
of
doing
a
demo
to
you
guys,
I
went
a
little
bit
over
the
top
and
created
a
little
bit
more
of
this
example.
B
B
So
that's
going
to
be
the
demo
of
the
day,
basically,
and
even
though
you're
probably
not
that
much
curious,
but
this
shows
how
the
the
connections
were
configured
down
here.
So
there's
a
plot,
a
bunch
of
plugin
arcs,
which
say
pipe
score
plugins
and
the
name
where
they
want
to
have
the
score
from
so,
for
instance,
there's
some
example
at
the
bottom.
It
wants
to
have
the
score
from
topology
scoring
and
the
invert
examples
of
these
guys
and
similarly
others
so
pretty
trivial
stuff
actually,
and
these
go
to
a
schedule.
B
A
profile
called
notebook,
topological
spread
scheduler
and
I've
disabled
a
few
plugins
just
because
there
would
be
too
much
prints
in
the
look
and
my
demo
is
just
about
lock
prints,
it's
not
going
to
be
overly
overly
fantasy,
but
let's
see
the
demo.
So
there's
the
port
and,
like
I
said
all
I'm
gonna
show
is
like
lock
prints,
so
I'm
gripping
through,
basically
the
scoring
and
the
scheduler
config
is
the
log
level
is
Max
to
spinal
tap
levels,
so
it'll
be
printing,
all
the
scores.
B
So,
let's
see
do
we
get
a
demo
effect
here?
No,
we
didn't
so,
let's
go
and
see
so
the
node
resources
fit
was
where
things
were
supposed
to
start,
so
it
gave
the
nodes
which,
by
the
way,
were
present
in
those
topologies
that
I
presented
earlier
this
kind
of
a
scores
so
they're
pretty
fairly
similar
to
each
other,
and
then
we
have
several
plugins
going
on
here.
B
So
we
have
the
topology
scoring
thing,
which
should
be
giving
the
best
score
to
the
one
topology
which
had
two
nodes
in
it,
and
it
is,
if
you
remember
the
names
of
those
they
were
dglh2
and
CML
S2.
So
since
node
resources
fit,
if
you
calculate
the
sum
of
those
that
topology
had
like
way
more
than
the
other
topologies,
which
only
had
one
node
in
them.
So
this
is
the
best
topology,
so
they're
getting
100
and
the
the
rest
of
the
topologies
only
had
a
single
node.
B
So
among
the
rest
of
them,
this
is
the
worst.
So
it's
getting
a
zero.
The
CFLs
and
the
rest
are
then
linearly
pretty
close
to
the
worst
because
they
only
had
one
node.
So
that
kind
of
like
proves
that
the
topology
scoring
works,
but
at
the
same
time
there's
this
composition
example
actually
running
so
the
code
that
I
was
showing
you
and
it
has
the
exact
same
score
as
the
topology
scoring.
B
So
yeah
no
surprises
there,
so
these
guys
exact
same
score
as
a
composition
so
functionally
it's
the
same
in
terms
of
you
know
getting
my
use
case
where
I
wanted
to
get,
but
conceptually
this
pipeline
is
totally
different
and
it's
worth
mentioning
Maybe,
the
other
plugins
as
well,
so
the
invert,
the
example.
What
is
that
so?
B
Basically,
the
invert
example
is
a
very
simple
scoring
plugin,
which
takes
the
input
score
from
basically
any
configured
other
scoring
plugin
and
calculates
a
100
minus
the
input,
so
it
inverts
the
logic
which
is
well,
maybe
not
so
useful,
as
just
you
know,
as
a
demo
here.
But
if
you
start
to
think
about
you
know
the
potential
of
such
a
thing,
you
could
invert
any
scoring
logic
with
that
kind
of
a
concept
which
could
be
handy,
although
it's
not
overly
useful.
B
If
you
have
the
same
weight
for
you,
know
the
input
and
this
guy,
because
they
will
be
canceling
each
other
out,
but
which
is
by
the
way
happening
in
this
demo,
but
anyway
putting
that
aside.
So
the
the
last
one
was
the
sum
example.
So
this
is
like,
if
I
show
the
graph
again.
So
what
was
the
topology
again?
B
So
no,
actually
it's
the
demo
is
probably
the
one.
So
this
is
summing.
The
invert
example
of
the
topology
scoring
okay.
So
those
were
the
configured
things,
so
it
should
be
summing.
The
inversion
and
the
topology
scoring
so,
for
instance,
CFLs
has
a
score
of
26
okay,
so
invert
has
26
and
the
topology
Series.
So
that
adds
up
correctly,
but
let's
see
something
CFL
S5
well,
that
was
17
and
the
topology
score
was
10.
B
So
again,
yeah
adds
up
correctly,
so
kind
of
like
proves
the
point
it's
calculating
correctly
and
that
kind
of
ends
the
demo
so
yeah.
These
were
the
let's
say:
plugin
we're
using
Alternatives
that
I
could
come
up
with
I
kind
of
like
the
the
pipelining
idea
myself,
because
it
has
lots
lots
of
more
potential
to
it.
But
unfortunately,
that's
of
course,
a
framework
site
change,
although
not
a
big
one.
I
only
spent
less
than
80
lines
of
code
there,
but
yeah
anyway.
It
is
what
it
is.
B
The
the
sources
by
the
way
are
available
at
my
Fork
of
of
the
scheduler
so
and
it's
actually
at
the
at
the
meeting
agenda,
there's
a
link
to
there.
So
if
anybody
is
interested
in
taking
a
look,
please
do
and
I
think
it's
time
for
questions.
Does
anybody
want
to
see
anything
like
from
the
sources?
Oh
I
could
maybe
show
what
the
one
of
those
let's
say-
the
invert
for
instance,
what.
A
A
B
Well,
it's
that's
a
good
question,
but
basically
what
I
started
with
was
you
know
the
desire
to
have
a
new
plugin,
because
we
have
a
use
case
for
that
now.
This
pipelining
thing
that
I
showed
you
know
it's
nice
to
have,
but
it's
I'm
not
going
to
die
if
I
don't
get
it,
but
there's
potential
in
that
one.
So
I
mean
it's
open
for
debate
whether
it
would
be
beneficial-
and
you
know,
ideas
are
welcome.
I'm
just.
A
C
So
I
joined
late,
but
I
guess
I
gave
the
idea
of
the
framework
changes
that
so
can
it
set
some
music
scenarios
that
some
users
want
to
do
some
run
time,
pretty
elastic
debugging
on
some
particular
path.
We
don't
know
the
rules.
We
cannot
set
that
statically
in
the
Upstream.
So
maybe
this
framework
can
be
useful
for
before
you
scanning
those
kind
of
mechanisms
in
there
like
their
input,
runtime
production,
environment
and
some
past
guests.
C
B
It's
not
not
a
thing
so
yeah,
it's
a
it's!
This
pipelining
idea
actually
gives
more
like
flexibility
to
add
things
after
another
plugin,
for
instance.
Now
we
have
this
pretty
much
hard-coded
normalization
functions
and
things
like
that.
You
could
take
this
further
and
you
know
pipeline
another
plugin
doing
normalization
functions
or
whatever
you
want
after
one
or
the
other
and
like
I,
said
this
inversion
logic
or
anything
that
can
be
quickly
calculated
and
alter
the
scores
of
one
Plugin
or
the
other.
B
B
Yeah
so
in
my
particular
case,
I
want
to
have
both
the
node
resource
of
feed,
resource
results
and
I
want
to
have
them
summed
into
the
actual
end
score
as
well,
and
I
want
to
have
this
topology
score
for
for
the
topologies,
so
I
want
to
have
them
both,
but
there
could
be
situations
where
you
know
you'd
like
to
tweak
the
results
from
some
plugin
with
some
kind
of
a
mathematic
function
like
I
was
explaining
earlier,
but
you
wouldn't
really
want
to
anymore.
Have
the
original
Source
plugin
score
as
such?
B
You
would
just
want
to
have
the
end
result
after
you
know
pipelining
it,
so
that
would
be
require
a
little
bit
more
work
instead
of
the
current
logic,
which
is
summing.
Everything
up
to
the
final
score
of
the
nodes.
You'll
probably
need
to
have
something
similar
to
the
summing
plugin.
That
I
was
showing
doing
things
instead
of
the
current
logic,
which
is
summing
everything
up
but
yeah.
It's
more
of
a
flexibility
thing
allowing
for
tweaking
the
scores.
C
C
B
B
On
it
was
based
on
basically
looking
at
the
arguments
of
the
the
plugins,
so
there
was
a
special
meeting
given
to
one
special
argument
of
the
plugins,
so
that's
the
way
that
it
works.
It's
it's
a
really
small
patch.
Actually,
so
it
won't
take
a
long
time
for
anybody
to
look
into
it's
less
than
80
lines,
and
that's
all
that
is
okay.
A
So,
for
once,
the
no
resources
fit
plugging
is
one
of
the
fastest
plugins
we
have
so
that
by
itself,
like
being
able
to
you
know,
reduce
calculation
might
not
be
as
important
both
depending
on
the
use
case
right
now,
if
you're
thinking
of
thousands
of
posts
per
second-
maybe
it
is
important,
but
yeah
back
back
to
my
original
point.
He,
if
you
see
value
on
having
this
Upstream,
it
would
be
good
to
gather
feedback
from
others
to
see
where
it
is
worth
to
to
Upstream.
B
A
A
Right
so
there
is
a
yeah.
There
is
a
repository
for
that
I.
Don't
have
it
at
hand
if
you
don't
find
it.
Let
me
know
in
this
slack
and
I
can
look
it
up,
but
yeah.
That
would
be
a
better
location
for
this
kind
of
content
on
his
I
guess
a
bit
more
free
form,
and
you
know
the
audience,
is
it
fits
better,
the
audience?
Okay,
so
you
could
consider
that
and.
A
A
B
It's
it's
kind
of
like
it
is
very
much
topology
related
and
this
latter
one
that
I
was
demonstrating.
It's
a
generic
topology
score,
so
basically
you
can
hook
it
up
with
anything.
It
doesn't
have
to
be
the
node
resources
fit
any
score.
Plugin
can
be
hooked
up
to
that
thing,
I'm,
not
sure
if
they
are
really
useful.
But
if
you
wanted
to
calculate
topology
scores
based
on
something
else
you
could.
This
won't
allows
it.
B
B
A
A
A
C
Yes,
I
was
but
Ernest
was
a
long
time
ago.
Yeah
have
I
tried
to
refreshing
my
memory
and
yeah.
The
beginning
says:
there's
nothing
to
significant
change.
Since
I
lost
a
review
there,
I
will
yeah
check
that
I.
Think
I
did
do
some
live
debugging
with
some
with
chain
on
the
internet,
but
I
just
need
to
refresh
the
memory.
I
will
take
a
look
yeah,
but
if
you
are
interested,
you
can
take
a
look
I
believe
yeah.
This
has
been
postponed
by
more
than
one
release,
so
yeah
it's
good
to
have
exercise
on
it.