►
From YouTube: 2022-05-26 Kubernetes SIG Scalability Meeting
Description
Agenda and meeting notes - https://docs.google.com/document/d/1hEpf25qifVWztaeZPFmjNiJvPo-5JX1z0LSvvVY5G2g/edit?ts=5d1e2a5b
A
Of
it,
so
this
is
six
callability
meeting
26
may
2020,
and
today
we
have
one
topic
on
our
agenda,
which
is
fake,
cuboid
manager.
A
C
Oh
yeah
yeah.
I
can
sorry
yes,
I,
like
it:
okay,
okay,
cool
cool,
so
today
I'm
gonna
to
present
a
tool
that
can
enable
simulating
a
fake
cluster
with
a
lot
of
fake
cube
nodes,
and
we
know
that
there
is
already
a
tool
called
qmark
or
called
hollow
node
in
the
upstream.
C
C
C
C
Yeah
this
is
based
on
the
123
release,
and
after
this
you
can
see
that
there's
20
notes
right
yeah.
This
is,
we
all
know
that
it
works
and
I
think-
and
although
I
haven't
been
closely
watched,
the
progress
of
the
cubemark
for
a
while-
but
I
think
maybe
it's
the
still
the
latest
status
of
how
cubemark
running
underneath
using
the
holo
binary
right
and
with
this
definitely
it
works
in
most
of
cases.
But
in
terms
of
scalability
I
mean
how
you
scale
from
a
20
nails
to
a
2000
notes.
C
So
basically
the
usage
will
just
increase
in
a
lot
of
new
manner.
You
can
see
that
on
my
local
desktop,
most
mod
takes
55
megabytes,
so
20
will
equals
one
gigabyte.
So
if
that
means
it's
basically
impossible
to
run
two
thousand
or
five
thousand
notes
fake
nails
on
my
local
cluster
so
and
also
definitely
I
can
try
to
schedule
a
part
here
and
you
can
see
that
the
case
is
that
hollow
node.
C
I
think
it's
basically
was
developed
from
the
skeleton
of
the
kubrick
code
base,
so
it
removed
any
unnecessary
things
or
marks
on
interface.
So
you
can
see
that
there
is
something
not
totally
stripped
off,
because
it's
a
virtual
thing,
so
we
don't
quite
need
something
like
dns
or
stuff.
So
I
think
that's
the
thing
that
it
takes
extra
memory
that
we
may
not
need
so
but
affectionately.
It
works
right.
C
C
And
I'm
going
to
introduce
a
project
called
fake
cubelet,
so
it
works
slightly
different
with
the
holo
node.
So
basically,
it's
more
a
controller,
a
manager
running
there
and
the
watch
on
any
node
event
pushed
to
aps
server
and
then
take
over
the
node
events
and
also
it
will
take
over
the
partial
responsibility
of
cube
blade
like
to.
C
Transfer
the
state
of
a
penny
part
with
the
spec
new
name
to
running
state.
So
basically
it
just
does
the
minimal
things
to
maintain
the
state
of
the
path
as
well
as
the
habit
of
the
node.
So
it's
just
that's
the
minimum
thing.
It
doesn't
do
any
like
image
or
any
node
particular
things.
So
it's
a
basic
controller
and
okay.
So
here
you
can
see
that
it
has
a
local
running
control,
plane.
C
You
can
see,
there's
no
note
here.
Next,
I'm
gonna
choose
started
the
fake
cube
link
so
that
the
name
is
a
little
confusing.
Maybe
we
can
rename
it
to
like
fake
cubelet
manager
or
take
over
something
like
that.
So
next
I'm
gonna
to
spin
up
1000,
fake
notes
here
and
using
this
tool.
So
you
can
hear
the
full
command.
C
C
C
Maybe
there's
some
capacity.
I
should
download.
C
And
you
can
see
the
1001
node
can
be
take
over,
so
you
can
dynamically
build
it
in
no
spec
fake,
no
spec
and
spin
out
a
cluster
easily.
So
that
is
why
I
want
to
present
this
to
seek
scalability,
because
these
two
can
potentially
enable
you
to
run
maybe
cube
mark
in
your
local
laptop
with
limited
memory,
and
also
you
can
use
basically
right
now.
C
My
scenario
is
to
try
out
the
scheduling
behavior
by
given
different
scheduling
strategies
like
tuning
the
weight
like
something
else,
but
it
can
be
extended
to
some
control,
plane,
performance,
latency,
scalability,
test
with
the
minimal
memory
footprint.
So
that
is
basically
I
want
to
bring
it
up
to
schedule,
scalability
and
also
I
present
it
to
scheduling
as
well,
but
I
think
it
most
likely
related
with
scalability.
C
E
I
think
the
question
is
like
how
far
we
are
from
like
the
load
that
is
generated
generated
by
like
this
cubelet,
comparing
to
like
real
clusters,
because
even
with
cubemark,
we
are
seeing
pretty
significant
difference
in
terms
of
like
the
load
that
the
hollow
cubelet
is
generating
on
the
control
plane,
and
I
think
we
are
like
the
main
focus
is
or
the
the
the
main
focus
was
always
really
around
like
api
servers,
so
like
the
more
we
fake,
the
less
load
or
lower
load
or
like
smaller
objects
or
whatever,
and
so
on
and
so
on,
we
are
getting
on
the
api
server.
E
So,
to
be
honest,
like
the
biggest
benefit
I
see
exactly
in
the
use
case
is
that
that
that
you
mentioned
like,
for
example,
scheduler
where
we,
if
we
want
to
figure
out
like
how
fast
the
scheduler
is
or
how
big
improvement
like
the
certain
change
in
the
algorithm
and
scheduler,
it's
it's
bringing
up.
E
Fast,
how
fast,
maybe
I
don't
know,
controller
manager
is
or
like
replica
asset
controller
or
whatever
like.
If,
if
this
fancy
new
improvement
there,
I
can
see
a
huge
difference,
a
huge
advantage
of
that
for,
like
those
those
kind
of
use
cases,
I
don't
think
we
will
be
able
to
replace
our
existing
tests
where,
like
the
biggest
problems,
are
usually
at
the
api
server
level.
So
I
think
I
would
treat
her
I'm
very
supportive
like
to
see
that
building
some
tooling
about
that.
E
But
I
think
it's
it's
more
additional
thing
than
the
replacement
of
anything
that
we
have
currently.
C
C
I
just
proposing
that
not
necessary
to
replace
those
whatever
so
in
my
current
scenarios,
so
this
gives
the
user
a
very
convenient
and
performant
way
to
craft
a
local
cluster
with
two
thousand
fake
node
or
whatever,
then,
based
on
that,
we
can
build
like
cube,
scheduler,
simulator
or
or
controller
manager,
performance,
tester,
atop
that
this
fake
cluster,
so
that
is
the
most
benefit
of
that.
So
right
now,
I'm
building
some
like
the
cube
scheduler
simulator
to
test
some
configuration.
That
is
why
I
think
this
tool
is
pretty
useful
and
they
want
to
present.
E
Yeah
so
sure
I
absolutely
agree,
as
I
mentioned,
I
think
like
scheduler
is
not
the
only
place,
I
think
like
stimulate
or
evaluating
certain
changes
of
like
pretty
much
any
master
controller.
It
would
be
useful,
like
daemon.
E
I
think
it
would
be
fairly
useful,
so
it's
not
just
scheduler,
I
think
it's
so
from
that
perspective,
I
think
it
would
be
fairly
useful
for
for
multiple
people.
I
think
it's
as
I
mentioned.
I
think
it's
not
a
replacement,
but
I
think.
E
A
useful
thing
to
be
able
to
run
it
to
test
certain
things
at
scale
on
your
local
laptop.
E
C
C
E
Pet
project
right
so
yeah,
I'm
I
and
I
think,
I'm
I'm
definitely
supportive
for
like
saying
that,
it's
a
it's
a
useful
thing.
So,
if,
if
you
need
any
help
with
convincing
anyone
that
it's
actually
useful,
I'm
happy
to
help
with
that
yeah.
I
also
yeah.
C
I
also
talked
to
paris
and
some
other
guys,
so
basically
I
can
definitely
promote
it
to
sponsored
by
six
scheduling,
but
according
to
the
essence
of
this
project,
is
small
in
scopes
related
with
six
scalability.
So
I'd
like
to
propose
it
to
be
sponsored
by
six
scheduling.
If
you
find
the
masala,
don't
object.
E
I
think
the
question
is
like
about
maintainability.
I
think
we
I
will,
I
will
see
them
will
yeah.
I
will
take
most
of
the
maintainability
of
this
one
species
yeah,
I'm
I'm
definitely
not
against.
I
think
we
will
not
have
capacity
to
really
help.
A
Yeah,
I
would
say
that
definitely
you
know
like
if,
if
we
find
some
like
really
cool
scenarios
for,
for
example,
for
scheduler,
then
I
would
be
totally
fine
with
you
know,
adding
to
our
our
set
of
benchmarks
like
new
benchmark
using
this
fake,
fake
cubot.
I
think
it
would
totally
make
sense,
because
right
now,
I
think
we
we
use
cubemark
but
cube
mark
in
this
case,
for
example,
for
scheduler
or
any
any
other
master
controller.
I
think
it's
it's
not
necessary
and
we
could
probably
use
use
fake
googlet
here.
A
C
C
And
I
think
the
first
thing
yeah,
that's
the
I
think.
That's
the
next
step.
We
after
we
get
it
promoted
to
a
sig
schedule,
sorry
to
the
kubernetes
sig
project,
and
if
you
don't
object,
I
think
I
can
make
it
co-sponsored
by
by
v2
6.
C
How
do
you
think
so
that
I
will
take
most
of
the
review,
releasing
the
regular
development
responsibilities?
C
C
Sounds
great
sounds
great,
yeah
yeah.
I
think
that's
pretty
much
all
I
want
to
present
and
also
right.
I
just
demoed
that
how
it
works
with
a
local
image,
but
it
has
the
potential
to
be
used
by
a
regular
user
which
doesn't
necessarily
need
to
touch
the
code
base
right.
So
basically,
another
scenario
is
that
I
just
create
a
node
cluster
and
just
deploy
it
as
a
path,
and
then
it
will
take
over
all
the
later
nails
created
onto
the
cluster.
So
basically
it's
more.
Then
it
gets
more.
C
It's
usage
to
regular
users
who
can
okay,
I'm
a
just
another
experienced
kubernetes
programmer,
I
just
a
regular
user.
I
can
use
this
tool
to
spin
up
a
local
cluster.
I
can
do
some
like
deployment
tests
whatever
and
like
replay
my
production
workload
to
my
local
environment
and
then
test
out
how
to
yeah
compose
or
beam
pack
or
whatever
kind
of
scenarios
before
you're
rolling
out
to
the
production
cluster
or
turn
into
some
kubernetes
expert
for
help.
So
I
can
play
with
that
locally.
So.
C
A
C
A
So
I'm
wondering
if
we
have
anything
else
on
the
agenda.
I
don't
think
so.
So
thank
you
for
for
meeting
and
thank
you
for
presentation
and
see
you
in
weeks.