►
From YouTube: 2016-07-14 Kubernetes SIG Scaling - Weekly Meeting
Description
Public meeting recording of the Kubernetes Scalability SIG.
Check comments for meeting chat log.
A
B
A
B
A
A
A
So
we
just
had
a
there,
just
a
one
admin
note:
oh
I
recorded
last
week's
meeting
and
we
didn't
really
have
a
process
for
how
we're
going
to
post
it.
So
I
was
not
going
to
Sarah
about
it
and
what
I'm
going
to
do
is
set
up
using
the
group
credentials
that
we
have
for
managing
zoom
we're
going
to
set
up
a
YouTube
channel
and
I'll
just
upload
these
and
create
you
channel
I'll
post
it
here.
A
So,
but
some
that's
some
busy
work
for
me
to
work
on
here
for
the
next
week,
but
was
one
thing
just
a
follow-up
from
us
last
week.
David
did
a
presentation
that
it
probably
dragged
on
a
little
long,
just
because
we
we
hadn't,
really
quite
polished
everything
up
about
to
do
prometheus,
based
sharing
age
information,
I.
Think
David's
planning
on
coming
back
with
a
more
succinct,
a
more
succinct
version
about
based
on
some
of
the
feedback
we
got
as
well.
A
So
that's
pro
tentatively
for
next
week,
maybe
the
week
after
and
then
we
have
a
new
person
on
our
team.
Who
was
just
that
still
sleeping?
Was
it
go
for
comp
overcoming
and
it's
a
Mike
Venezia's
was
the
Cooper
Nettie's?
Can
the
Cooper
Nettie's
lead
at
viacom,
so
he's
joined
the
samsung
team
and
we're
back
to
I
was
thought
I
saw
I'm
cho
em
up
on
the
unbeaten
delist
linear,
a
true
we're
kind
of
renewing
some
work
we
were
doing
around
at
CD.
Performance
testing
is
multiple
configurations.
Primarily.
B
B
B
Did
have
a
couple
of
topics
but
America's
on
so
did
you
have?
Is
there
a
numbers
that
you're
aware
of
or
node
performance
in
scale
at
the
number
of
pods
per
node?
That
issue
is
kind
of
like
dragging
on
because
it
will
affect
density.
The
one
where
pods
per
node
total
number
of
pods
are
positive
or
total
number
of
parts
of
the
known.
C
Well,
I,
don't
know
if
it's
tested
anywhere
like
I
know
they
tests
like
on
smaller
machines.
It
should
be
okay,
I
did
I
think
they
test
that
I
put
100
or
200
something
but
like
when
you
are
moving
to
like
32
cosmos
in
chains
with
which
should
support
320
or
something
then
I,
don't
think
anyone
ever
tested
it.
D
B
C
E
A
I'm
I'm
going
to
turn
back
and
make
more
notes
here.
B
D
B
F
So
I
I
talked
about
a
thing
in
a
6k.
We
think
and
we
have
devised
little
discussion
with
I'm
Lana
and
it
is
sounds
like
he's
just
not
ready
for
attitude.
D
But
it's
like
significantly
better
and
in
particular
I
will
also
like
I'm
generally
testing
like
higher
throughput
in
like
Al
Gore,
2000,
no
casters
or
the
whole
system
like
whole
system
like
and
to
end
system,
with
not
many
controllers,
like
basically
only
replication
controller,
and
it
seems
that
we
are
ready
to
come
clean
crease.
The
QPS
limits,
like
obviously
the
best
case.
The
best
thing
that
we
would
be
able
to
do
is
to
have
some
back
pressure,
but
that's
something
that
is
not
going
to
happen
like
soon,
at
least
in
balloon,
1s
bandwidth
that
that.
B
Was
something
that
was
on
my
to-do
list
to
start
looking
into
so
I
once
I
get
rid
of
my
ridiculously
large
backlog
of
emails,
also
by
which,
if
we
can
start
segregating
or
labeling
a
little
better,
that
will
be
awesome
sauce,
because,
right
now
I
have
a
drink
from
the
firehose
of
cooper,
Nettie's
emails
to
regularly
I,
don't
know
that
other
people
also
feel
that
burn
so
to
speak.
Well,.
E
B
Guess
the
only
other
thing
I
can
report
out
is
that
we
were
doing.
It
probably
doesn't
affect
anyone
who's,
not
using
device
mapper.
But
it's
something
that
we
noticed
when
we
started
going
really
high
on
the
nodes,
high
densities,
so
we're
doing
densities
above
100
we're
doing
densities
at
250,
and
when
we
do
that,
we
see
the
memory
utilization
of
the
node
start
to
grow
really
high
and
we've
already
reported
it
to
the
appropriate
people.
B
B
No,
it's
see
advisor
Oh,
Steve,
ledger
hilton,
the
couplet,
so
the
memory
utilization
for
our
16
gig
machine
with
250
pods
doing
nothing
was
5
gigabytes
and
memories
just
to
pretty
much
track.
Those
pods
and
most
of
that
memory
was
spent
in
the
heap
profiler.
Most
of
the
memory
was
spent
in
a
disk
monitoring.
E
E
Somewhere
in
catching
up
on
my
backlog
of
tickets,
I
think
I
saw
like
a
sea
advisor
roadmap
ticket
living
out
there
somewhere.
Maybe
that's
that's,
got
some
information
on
this
issue.
E
B
D
D
A
D
B
D
B
D
D
Like
definitely,
we
can't
go
betta
betta
if
we
won't
improve
performance
but
I
as
a
as
far
as
I
know.
It's
not
the
only
thing
that
we
should
do
to
go
to
like
improve
to
beat
that.
So
I'm
not
sure
if
it
like
the
plan
is
that
it
should
go
to
beta.
But
I
don't
know
if
anyone
is
doing
anything
else
than
improving
performance
right.
E
This
is
one
of
those
questions
were
like
I
guess.
This
had
been
started
before
the
whole
features.
Repo
and
I'm
still
unclear
whether
that's
really
the
marching
order
the
community
is
taking,
but
I'm
still
confused
about.
There's
a
really
clear
document
about
API
is
moving
from
alpha
to
beta
2
GA.
It's
unclear
to
me
like
what
other
gates
or
criteria
are
needed,
and
so
it
sounds
like
if
Google's
really
got
is
really
interested
in
seeing
affinity
on
by
default.
D
E
No
I
agree
I
sort
of
like,
in
my
my
sweep
of
my
ridiculously
large
backlog.
I.
Do
you
remember
running
into
this
stuff,
but
I,
don't
recall,
seeing
an
umbrella
issue
that
ties
this
all
together.
That's
sort
of
I
guess
what
I'm
asking
for,
or
some
kind
of
document
that,
just
as
where
the
performance
criteria
are
listed,
it.
D
E
A
Hey
could
I
could
I
try
to
spit
back
when
I'm
hearing
here
and
see
if
I
got
it
right
wojtek.
So
the
affinity,
anti
affinity
feature
causes
pretty
big
hit
performance
you're
working
on
throughput
improvements.
The
goal
for
14
is
to
get
the
affinity,
anti
affinity
features
into
14,
but
not
to
have
performance
go
backwards.
Yes,
okay,.
C
C
Yeah
I
have
one
thing
for
like
a
like
a
well
a
question
for
everyone
like
I,
will
be
working
contesting
controllers
city
soon
and
or
controllers,
which
means
I
do
pretty
much
features
that
we
have.
I
like
and
and
all
the
sounds
like,
whatever
number
of
parts
in
a
service
or
something
so
a
type
of
skating.
E
E
I
thought
this
is
why
I'm
curious,
like
density
for
example,
involves
the
use
of
replication
controller
controller.
You
could
say
that,
like
the
time
to
ask
it
to
say:
please
create
300
pots,
and
then
you
see
300
pots
read
it.
You
don't
care
whether
they
went
running
it.
Ready
right
is
that
sort
of
an
example
than
SLO
for
a
controller?
That's.
C
Yeah,
that's
an
example
for
that
I
like
the
internal
solo
power
controller,
but
like
if
you
but
I,
also
interested
in
like
that,
the
SLO
for
the
world
for
the
time,
but
it's
also,
I
also
knee
into
type
to
define
like
what
does
it
mean
like
what
is
take
the
system
so
like
if
it's
the
first
control
replication
controller
or
do
you
want
in
the
yeah?
I
can
so
like
not
only
DD
like
response
times
and
throughput
and
all
the
other
like
normal
stuff,
but
I
also
the
load
in
the
environmental
which
we
are
testing.
C
C
B
B
C
B
Anything
well
I
mean
this
because
we
have
this
separate
because
in
Cooper
Nettie's
we
have
segregated
the
controller
from
the
scheduler,
which
is
different
from
other
systems
where
they
smash
them.
Together,
the
we're
always
going
to
have
this
weird
round
trip
back
and
forth
questions
so
like
we
can
I'm
happy
to
collaborate
on
a
document.
I
guess
the
question
will
be
like.
C
E
Feel
free
to
this
I
can
think
of
like
there's
some
action
they
should
take
within
some
latency
window.
If
you
could
probably
describe
each
controllers
responsibility
in
terms
of
do
one
thing
and
then
do
it
within
a
certain
amount
of
time
that
gives
us
some
leather
upper
bound
on
the
system.
Performance,
yep.
C
E
In
terms
of
like
the
actual
constructs
that
are
interesting,
I
might
start
with
whatever
constructs
the
official
adults
are
now
scheduled
with
and
I
legitimately.
Don't
know
whether
that's
replication
controllers
replica
sets
deployments
or
pet
sets,
but
it's
it's
migrated
from
those
from
release
to
release.
That
would
probably
be
a
good
place
to
make
sure
you've
got
guarantees.
First.
E
I
think
there's
something
about
that.
The
end-to-end
tests
I
think
installed
a
été
image
polar
that
should
preload
every
note
all
the
images
necessary.
So
if
you
could
say
that
you
know,
given
certain
conditions
with
little
nodes
pre-loaded,
you
can
try
to
take
out
some
of
the
environmental
conditions,
cool.
D
C
E
Yeah
totally
fair
and
just
sounds
like
you
want
to
remove
environmental
conditions
et
image,
polar
and
mandating.
Some
prerequisites
on
the
cluster
state
under
test
is
first
and
then,
since
I'm
sort
of
under
the
impression
that
eventually,
the
controller
manager
is
going
to
turn
into
this
thing,
where
each
of
the
controllers
start
to
float
into
their
own
positive
process
these
elsewhere
within
the
cluster.
It
makes
more
sense
to
be
to
find
the
controllers
that
are
most
meaningful
to
users,
and
so
the
first
thing
you're
going
to
encounter
the
add-ons
their
scheduled
into
their
cluster
right.
C
G
So
I,
just
one
more
question
about
so
currently
what
I
see
in
the
co
is
that
I
feel
funny.
Look
isn't
cash
in
EK,
so
it's
like
antique
plates,
something
called
undecorated
like
anyway
like
is
it
any
any
reason
why
it's
not
catching
you
guys
ever
so.
If
I
have
a
lot
of
party
resources
like
I
like
if
I
do
a
less
or
something
like
it's,
it's
gonna
call
from
a
CD
/,
most
knowledge
jointly.
That's
gonna
be
consuming
I.
D
A
Well,
we're
we're
out
of
time
for
today,
good
discussion.
I
was
busy
trying
to
make
notes,
but
some
of
my
notes
are
a
little
bit
half
done
here.
I
see
some
folks
in
trying
to
help
them
get
up-to-date.
So
just
a
quick
call
to
take
a
look
at
the
notes
and
help
get
them
updated
and
I
think
we're
done
for
today.