►
From YouTube: Kubernetes SIG Node 20230718
Description
SIG Node weekly meeting. Agenda and notes: https://docs.google.com/document/d/1Ne57gvidMEWXR70OxxnRkYquAoMpt56o75oZtg-OeBg/edit#heading=h.adoto8roitwq
GMT20230718-170525_Recording_1920x880.mp4
A
And
hello,
hello,
it's
July,
18
2023!
It's
a
signature,
weekly
meeting
welcome
everybody.
Today
we
have
a
couple
items
on
agenda
and
the
biggest
item
is
I
mean
outside
the
agenda.
We
have
a
court
list
today,
so
if
you
have
anything
that
you
want
to
be
merged,
please
on
a
slack
or
just
make
some
noise
to
make
sure
that
it's
been
looked
at.
We
don't
have
too
much
of
bandwidth
for
our
Brokers,
but
I
think
we
can
have
some
okay.
First
item
is
Dixie.
Do
you
want
to
speak
about
it?
B
So
we
were
planning
to
do
beta
for
memory
Qs
in
128,
but
I,
don't
think
we'll
be
able
to
so
I
Ran
So.
The
essential
thing
that
memory
Qs
does
is
it
adds
memory
dot.
It
adds
memory
throttling
by
setting
memory,
dot
high
value
for
in
the
new
C
group.
B
V2
controller
and
I
just
tried
to
run
a
very
simple
test
without
kubernetes
I
created
a
container
and
while
it
was
sleeping,
this
container
basically
adds
one
megabyte
of
memory
in
every
run
of
the
process,
and
while
it
is
sleeping
here,
I
change
the
memory
Max
and
memory
high
value
for
it.
Like
I,
said
the
memory
high
as
around
19
megabytes
and
memory.
Max
has
20
and
what
I
noticed
was
around
19.
B
B
So
in
this
scenario
we
see
that
there
is
just
no
signal
coming
from
the
container
to
do
any
action
there.
It's
just
stuck
there,
so
I'm,
not
very
sure
whether
we
will
be
able
to
add
this
feature
so
I'm
right
now
talking
to
some
internal
Corner
teams,
trying
to
see
what's
happening
there
and
also
trying
to
understand
how
the
reclaim
works
I
think
there
is.
There
are
a
few
signals
that
we
can
read
in
memory:
dot,
stat
file,
PG
scan
and
PG
steel.
When
I'm
reading
those
values.
B
I
do
see
memory
reclaim
happening,
but
there
is
some
deadlock
scenario.
That's
going
on
so
I'll
have
a
I'll,
have
a
postmortem
sort
of
a
Thing
Once
I
am
able
to
understand,
what's
happening
and
then
decide
whether
we
want
to
proceed
or
not
thanks.
One
question
I
have
dikshita
is:
did
we
look
at
the
PSI
and
get
anything
useful
there?
B
B
All
right,
I'll
share
it
with
our
kernel
folks
as
well
yeah.
That
will
be
really
great,
because
I
would
need
help
in
understanding
the
kernel
code
also
as
to
how
memory.I
works,
where
you
are
from
just
using
this
test,
3,
which
is
running
without
even
kids.
It's
just
using
a
simple
Docker
container
and
setting
the
values.
So
if
you
could
run
this
test
and
critically
analyze
what's
happening
here,
it
will
be
really
helpful.
B
C
D
Yeah
I,
just
I
just
wanted
to
add.
Also
just
really
awesome.
I
think
we
caught
this
before
we
promoted
to
Beta,
because
it's
I
think
critical
thing
that
you
know,
containers
are
still
expected
to
kind
of
you
know
be
able
to
to
right
if
they
go
over
the
limit,
so
I
really
just
want
to
think
think
Dixie
for
catching
this
before
we
pronounce
the
beta.
So
we
can
further
investigate
what's
going
on
and
you
know
figure
out
what
the
plan
is
so
excellent.
Yeah.
E
Hi,
hello
yeah:
this
is
Joshi
and
I
have
two
colleagues
on
the
meeting
as
well,
and
so
we,
we
are
all
from
by
dance
the
parent
company
of
tick
tock
yeah.
Thanks
for
those
awesome
features,
we
leverage
the
CPU
manager
and
the
In-Place
vpa
to
build
our
internal
service
product
and
I.
E
Think
we
meet
a
lot
of
issues
and
also
accumulate
some
experiences
in
Downstream,
especially,
is
the
new
CPU
static
policy,
and
also
we
make
the
CPU
Manager
work
with
the
In-Place
vpa
feature
and
also
resolve
some
performance
issues
of
the
In-Place
BPA
and
their
different.
Like
Corner
cases
now,
I
think
we'd
like
to
try
to
discuss
with
the
community
and
also
try
to
seek
an
opportunity
to
Upstream
our
changes.
So
we
would
like
to
definitely
like
to
get
more
familiar
with
the
process.
E
C
I
would
I
would
rather
State
a
problem
what
kind
of
the
performance
problem?
What
kind
of
the
the
problem
you
want
to
address
right,
so
I
think
the
Francis
already
share
some
his
opinion.
I
agree
with
him.
Maybe
start
if
you
want,
because
the
problems
statement
maybe
start
on
the
Google
Doc.
So
people
can
comment
and
ask
a
question
and
then
then
we
can
go
with
the
cap
and
the
pr
yeah.
E
That
sounds
great
yes,
I
can
briefly
share
the
problems
we
made.
The
first
problem
is
the
the
core
mining
issues
and,
based
on
our
experiment,
we
noticed
if
we
can
schedule
the
hyper
thread,
of
course,
like
physical
course
that's
much
more
efficient
to
us
and
because
most
of
the
we're
building
serverless
products
most
of
the
like
instances,
they
are
not
always
busy.
So
that
means
we.
If
we
subscribe
to
other
like
physical
course,
the
the
data
shows
like
30
of
the
performance
Improvement
for
our
database
like
instances.
E
So
that's
the
that's
the
one
like
as
the
policy
we
like
to
add
I
noticed
the
community
also
have
the
like
the
distributed
distributor.
The
CPUs
across
normal
pneuma
is
a
little
bit
different
from
our
policy.
We
can,
we
are
allowed
to
cross
normal
or
not,
but
the
the
key
idea
is
to
make
sure
they
are
not
located
in
the
same
physical
core.
So
that's
a
that's.
E
A
static
policy
and
the
the
vpa
part
I
think
that
that
is
the
not
that
mature
and
it
doesn't
work
with
the
CPU
manager
when
we
do
the
scouting
and
we
make
that
part
our
works
and
also
make
sure
the
the
the
scaling
like
efficiency
and
latency
are
optimized
So.
Currently
we
miss
some
problems
that,
under
certain
conditions,
it
takes
like
up
to
a
minute
to
finish
the
vpa,
because
the
C
group
updates.
E
Sometimes
it
needs
to
wait
to
the
next
couplet
reconcile
Loop
to
pick
up
the
changes,
yeah
I
think
that's
the
major
like
changes.
I
think
we
did
in
the
in
the
downstream
yeah.
C
Is
renewable
request
and
can
you
put
it
in
the
dark
and
also
how
to
enhance
so
one
thing
I
want
to
pre-warning.
It
is
just
because
open
source
Community
try
to
provide
the
generic
Solutions,
so
we
cannot
promise
right
so
so
your
problem
statement
actually
I
heard
even
like
the
some
now
I'm
representing
the
gke
and
the
internal
gke
user.
Some
of
the
also
requests
the
similar
things
so.
A
C
That's
why
it's
not
a
surprise
to
me
so
back
to
the
open
source,
what
has
open
source
ahead,
but
but
do
we
try
to
abstract
right
based
on
the
customer
user?
Not
customer,
okay
user
thinks
it's
community
and
the
user
request
the
demand.
So
we
try
to
Upper
strike
those
kind
of
things
and
provide
the
solutions.
C
So
please
put
me
to
the
problem
statement
and
because
those
requests
are
being
heard
from
the
internal
customer,
this
is
from
all
site
that
I
first
time,
I
heard
I
haven't
seen
any
requests
like
this
before
so
put
it
into
the
dark,
and
we
try
to
abstract
those
kind
of
things
and
then
see
how
open
source
can
provide
the
solution.
So
yeah.
E
A
Okay,
so
we
have
next
steps
for
this
item
and
we
are
at
the
end
of
our
agenda,
so
I
think
we
can
give
her
the
rest
of
the
meeting
time
back
for
reviews.
If
there
are
anything
else,
please
speak
up
right
now.