►
From YouTube: 2016-09-01 Kubernetes SIG Scaling - Weekly Meeting
Description
2016-09-01 Kubernetes SIG Scaling - Weekly Meeting
A
A
B
C
A
B
I've
been
trying
to
dig
up
issues
the
link
that
I
step
placed
in
the
agenda
and
maybe
I'll
just
extra
actively
I
had
some
quick
questions
on
two
things:
I've
noticed
in
pull
requests
one.
There
was
a
pull
request.
We
talked
about
here
or
an
issue
about
increased
API
server
memory
usage
and
at
first
we
were
thinking
that
this
was
tied
to
the
garbage
collector
I.
B
Think
the
last
thing
that
I
saw
was
we
decided
to
close
that
out
because
it
seemed
like
it
was
dependent
upon
the
order
in
which
density
tests
were
run
versus
load
tests.
I
was
kind
of
I,
don't
know.
I
kinda
wanted
to
riff
on
that
for
a
little
bit,
because
I'd
sound
like
there's
some
sort
of
increased
memory,
utilization
that
happens
period
and
maybe
the
tests
are
checking
for
too
tight
abound.
B
If
you
order
in
which
were
running
at
tests
is
causing
us
problems
or,
conversely,
the
tests
are
leaving
behind
way
too
much
memory
utilization,
and
that
also
seems
like
something
worth
chasing
down.
I,
just
wasn't
like
I
seem
like
a
she
was
clothed
like
they
don't
have
to
worry
about
it
for
release,
but
I
wasn't
sure
if
we
were
going
to
follow
up
on
that.
So
that
was
one
thing.
The
second
thing
was
just.
B
D
D
I
would
like
to
look
into
after
the
release
if
anyone
can
help
with
that,
that
would
be
well
so
great
answering
the
second
one
we
had
some
like
Q
mark
was
broken
few
times
recently,
so
it
was
like
just
failing
it
wasn't
starting
I,
not
aware
of
any
like
in
a
healthy
situation,
assuming
that
it's
not
broken
I'm,
not
aware
of
like
significantly
higher
error
right
on
starting
the
classroom.
So
if,
if
you
have
any
like
links
to
it
or
like
links
to
PRS
words
or
like
oh.
B
Yeah,
like
I,
said,
I'll
see
what
I
could
pick
up.
I
just
know
that
Jeff
Grafton,
like
there
was
some
testing
for
changes
put
into
triage
off
failures
to
cluster
fail
to
come
up
as
opposed
to
whatever
the
test
was
spitting
out
about
it,
and
this
was
coming
up
across
a
number
of
issues.
It
just
seemed
like
to
mark
was
one
of
the
two
mark.
Five,
I
think,
was
one
of
the
ones
that
frequently
and.
D
B
D
B
D
A
E
Do
we
actually
have
a
lot
of
data?
We
finished
off
a
thousand
notes?
Well,
unsuccessfully,
but
we
we
were
able
to,
and
we've
logged
some
issues
against
it,
like
Edith
out
I,
think
it's
worthwhile
to
maybe
like
overview
what
we
do
for
some
of
our
experiments,
because
we
we
don't
necessarily
use
all
of
up
streams,
testing
facilities,
we
use
parts
of
it,
but
we
also
have
our
own
test
jig.
E
So
when
there's
it
up,
there's
a
repo
that's
available
that
we
use
called
the
svt
repo,
it's
open
shift,
SVT
and
in
there
we
basically
have
an
open
shift,
scalability
sub-project
and
underneath
our
there's
a
declarative
way
that
we
use
for
loading
or
cluster,
where
we
specify
very
simple,
like
we
want
10,000
services,
6,000
projects,
these
textiles
and
other
different
resource
objects
right
and
as
long
as
we
defined
like
the
the
boundaries
and
the
templates
for
those,
it
goes
ahead
and
increase
those
things
for
us
in
a
in
a
stepwise
increased
way.
All
right.
E
So
a
lot
of
times
will
exercise
using
some
of
the
templates
will
exercise
things
pretty
harshly
it
much
worse
than
density
is
so
we
might
see
things
that
are
kind
of
weird.
One
of
the
things
we
saw
was
the
scheduler
I
was
going
to
eat
up
a
ton
of
CPU,
and
we
love
that
issued
what
took
took
a
look
at
it
too.
E
D
Yeah
I
wrote
on
this
issue
that,
like
it
was
like
service
affinities
predicate
for
priority.
I'm
not
sure
like
I
look
into
it
and
the
problem
is
that
or
actually
maybe
another
problem
like
or
yeah
might
be
a
problem.
So
basically
the
problem
is
that
we
are
not.
We
don't
have
like
how
like
Reggie
started
by
default.
I
freakin
in
Google.
That's
why
we
probably
missed
during
/,
optimal
Isaac
optimizations
that
we
did
it
but
sure
I
will
look
into
it,
but
not
for
1.4.
Unfortunately,
oh
yeah.
E
Open
shift
work,
but
it's
generally
applicable,
so
other
people
could
use
or
rejigged
some
of
the
things
if
they
wanted
to
to
use
it
against
the
community
work.
So
a
lot
of
it's
been
written
with
the
multi-tenant
environment
in
mind.
So,
instead
of
us
like
directly
opening
our
tests
are
really
smart,
right
or
they're
smarter,
I
would
say
really
smart,
they're,
smarter
and
the
fact
that
they
only
open
a
single
client
connection.
Then
they
do
a
bunch
of
work
at
a
close,
the
kind
of
connection
in
our
tests.
E
E
E
D
Yeah,
this
is
something
I
also
so
so
I
send
some
PR
at
some
point
when
there
wasn't
like
very
simple,
like
four
lines,
we
are
because,
like
a
bunch
of
times,
but
Spencer
is
basically
on
validation
of
data
that
we
are
reading
from
at
CB
and
my
opinion
doesn't
make
much
sense
to
recreate
data
that
we
already
validated
with
writing
them,
but
yeah
yeah
I.
What.
E
I
had
that
covered
with
Daniel
a
while
ago,
we're
like
at
some
point.
We
even
do
the
decode
like
I,
don't
even
know
why
we
even
decode,
sometimes
because
we
decode
and
validate,
and
that
alone
is
a
lot
of
extra
work,
seeing
how
we
did
all
the
effort
to
encode
it
to
write
it.
So
it's
like
we're
the
writer
but
we're
checking
what
we
read.
Even
though
it's
going
to
be
slapped
across
the
wire
Dakota.
Yet
again,
so.
D
E
Yeah
I
totally
agree.
As
I
mentioned
the
heat
vs.
RSS.
We
need
to
look
into
that
to
figure
out
if
there's
a
way
to
trigger
or
force
GC,
because
sometimes
we're
the
in
memory
usage
is
super
high
by
comparison
to
the
heap
usage.
So
like
we
have
as
part
of
our
testing,
we
run
svt
in
a
on
a
run.
Then
we
use
p
prof
or
we
use
p
bench
to
periodically
pull
p
prof
data
and
in
the
heap
p
prof
data.
A
Think
so
far
our
are
111
to
based
work
seems
to
be
the
right
path,
although
we
haven't
put
enough
work
into
that
just
yet
to
really
be
confident
on
that
I'm.
Just
following
up
on
notes
from
last
week,
we
are
starting
to
do
more
work
on
the
config
on
the
config
dump.
In
fact,
we
were
gonna
get
some
time
with
Joe
sometime
soon
to
kind
of
walk
through
some
of
our
planning
on
that
I'm.
A
Not
really
sure
this
is
the
right
sig
for
us
to
be
kind
of
hosting
this
discussion,
but
I
don't
mean
it's
something.
That's
important
to
us
and
has
been
kind
of
you
know
important
in
this
in
this
meeting.
So
I
think
maybe
we'll
we'll
comment
it
on
on
it
here
until
it
seems
obvious,
we
should
move
it
elsewhere.
So.
C
C
Is
it
signode,
but
I
mean,
like
there's
other
stuff,
that
we're
going
to
want
to
configure
also,
like
the
you
know,
through
config
maps,
also
yeah,
but
that
particular
one
about.
C
All
right,
so
it's
probably
worth
at
least
pulling
together
interested
parties
there
once
we
have
maybe
a
strawman
of
how
we
want
to
dump
this
stuff.
It's
kind
of
like
you
know
everything
we
do
cuts
across
everything
and
so
I
hate
sending
these
like.
Hey
we're
doing
this
crab
to
to
Cooper
Nettie's
dev.
But
I
think
this
is
another
one
of
those
right,
but
this
all
just
does
relate,
or
you
know,
planning
for
things
to
track
for
1.5.
That
cycles
getting
started.
C
E
It
general
we
have,
you
know
we
have
a
data
model,
/
throughput
problem,
we've
never
really
addressed.
We've
we've
we've
handled
a
lot
of
the
quote-unquote
bigger
ring
like
getting
out
of
her
own
way
of
getting
the
cluster
larger,
but
we
haven't
really
addressed
speed
issues
and
concerns,
and
also
address
some
like
as
wojtek
was-
and
I
were
talking
about
earlier-
address
some
of
the
data
model
problems
of
the
system,
which
I
think
is
let
the
next
step
of
scale.
If
you
want
to
first
to
get
it
big,
then
you
get
it
fast
right.
I.
E
C
A
C
I
just
submitted
a
PR
and
I
I
copied
this
egg
on
it,
so
you
guys
can
can
look
at
that
for
the
for
the
community
repo
so
excellent.
I
did
my
manual
google
docs
to
mark
down
translation.
I
started
looking
at
tools
and
it's
one
of
those
things
where
you
can
spend
more
time,
picking
and
fussing
with
the
tools
than
to
just
go
ahead
and
do
it.
So
I
just
today
see
it's
a
story
of
every
diagram.
I've
ever
created.
C
B
D
So
my
understanding
is
that
there
should
be
betta
cut
release
like
later
this
week,
which
means
like
today
or
tomorrow,
I
think
but
we'll
see
if
that
happens.
So
it's
definitely
much
better
than
it
was
even
a
week
before,
like
in
terms
of
all
those
flags
and
stuff
like
that
and
I
think
that,
from
the
feature
perspective,
we
are
done.
D
D
B
No,
that's
that's
good
to
hear
that's
sort
of
what
I
was
seeing.
I
just
he'd
be
cool
if
the
beta
was
out
today,
because
a
bunch
of
people
are
gonna,
start
peeling
out
for
vacation
on
the
USC
line,
so
I
think
like
we
may
all
come
back
and
try
to
remember
wait
what
data
on
tuesday
or
wednesday,
so
it
great
to
get
some
time
with
that
before
them.